A large-scale knowledge graph construction method for the financial field
By constructing a common sense knowledge base in the financial field using convolutional deep neural networks and reinforcement learning, and by utilizing graph representation learning modules and distributed computing engines, the heterogeneity and independence issues of large-scale financial knowledge graphs are solved, achieving efficient knowledge integration and querying, and improving data utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- RES INST OF YIBIN UNIV OF ELECTRONIC SCI & TECH
- Filing Date
- 2023-09-20
- Publication Date
- 2026-05-29
AI Technical Summary
The construction of large-scale knowledge graphs in the financial field suffers from heterogeneity and independence issues, which makes it difficult to integrate diversified data and perform efficient reasoning and computation, resulting in insufficient knowledge quality and timeliness.
We employ convolutional deep neural networks for entity recognition and relation extraction, combine end-to-end knowledge base question answering and reasoning methods with reinforcement learning, construct a commonsense knowledge base, perform joint representation learning through a graph representation learning module, and utilize Spark and Flink to build a large-scale computing engine for real-time computation.
It achieves efficient integration and knowledge sharing of multi-source heterogeneous data, improves the efficiency of knowledge query and reasoning, reduces the workload of manual processing, has scalability and adaptability, and supports applications such as user panoramic profiling and risk assessment.
Smart Images
Figure CN117273128B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of knowledge graph technology, and more specifically to a method for constructing large-scale knowledge graphs for the financial sector. Background Technology
[0002] Large-scale knowledge graph construction technology has significant application value in the financial field. It can integrate massive amounts of financial data and knowledge into a structured graph, helping financial institutions better understand and utilize data to support decision-making and risk management. Key issues that urgently need to be addressed in the construction of large-scale knowledge graphs in the financial sector include: representation and automatic acquisition of multiple types of knowledge; integration of multi-source heterogeneous data in the financial field; efficient reasoning and computation of massive amounts of knowledge; and improving the quality and timeliness of knowledge computation. Summary of the Invention
[0003] To address the problems existing in the prior art, this invention provides a method for constructing large-scale knowledge graphs in the financial field, aiming to solve the problem of constructing diversified large-scale knowledge graphs caused by the heterogeneity and independence of large-scale domain knowledge.
[0004] A method for constructing large-scale knowledge graphs in the financial sector, comprising:
[0005] S1: Obtain financial and economic data from open knowledge bases, perform concept extraction and knowledge extraction, and then construct a large-scale knowledge graph;
[0006] S2: Construct a commonsense knowledge base based on a large-scale knowledge graph;
[0007] S3: Expand the knowledge base of common sense using large-scale knowledge graphs;
[0008] S4: Utilize the graph representation learning module for joint representation learning of common sense knowledge base and large-scale knowledge graph.
[0009] Preferably, the knowledge extraction in S1 includes entity recognition and relation extraction, and the knowledge extraction process includes:
[0010] We employ a deep convolutional neural network to address the text semantic representation problem in the knowledge acquisition process, reducing the excessive reliance on natural language processing tools and the error accumulation problem of traditional methods.
[0011] An end-to-end knowledge base question answering and reasoning method based on reinforcement learning is adopted to transform the traditional symbol-representation-based question answering and reasoning process into a learnable problem based on semantic numerical computation, thereby performing entity correlation calculation and knowledge reasoning.
[0012] Preferably, the common sense knowledge base in S2 includes an instance base, a rule base, and an association base, wherein the instance base includes common sense instance representation and common sense knowledge acquisition; the rule base includes constructing a financial domain lexical ontology and formulating rules; and the association base includes calculating a target similarity matrix.
[0013] Preferably, the knowledge expansion described in S3 includes forward reasoning and backward reasoning. Forward reasoning uses existing facts in the knowledge graph, such as relation triples, to match the premises of a rule. Once the premises of the rule are satisfied, the facts of its conclusion are added to the common sense knowledge base for expansion. Backward reasoning first assumes a target, then finds a rule whose conclusion matches the target, and then recursively proves that the premises of the rule are all true. Once the proof is true, the facts of its conclusion are added to the common sense knowledge base.
[0014] Preferably, the graph representation learning module includes a transformation layer, a model layer, and an interface layer. The transformation layer transforms the original knowledge graph into a standard graph and obtains its adjacency matrix representation, while providing a lookup table for easy querying of vertex representations. The model layer takes the adjacency matrix as input and obtains a vector representation matrix composed of low-dimensional vector representations of each vertex through the graph representation learning model. The interface layer obtains the vector representation of the classification label through the lookup table according to the actual task requirements, transforms it into a probability distribution, and further transforms it into a graph association matrix as a bias of the classification result of the deep learning module to indicate the learning process and accelerate convergence.
[0015] Preferably, the transformation layer is used to preprocess the knowledge graph into a standard graph to facilitate subsequent input into the graph representation learning model, including: given a knowledge graph (H,R,T), where H is the head entity set, R is the tail entity set, T is the relation set, and all predicates involved in the knowledge reasoning rules constitute a set U.
[0016] Preferably, the model layer includes a graph representation learning model. The process of applying the graph representation learning model includes: first, reconstructing the neighborhood structure of each vertex to preserve local characteristics through unsupervised learning; second, using Laplace mapping to learn the global characteristics of the graph further by using first-order similarity as supervision information through supervised learning.
[0017] Preferably, the interface layer transforms the graph representation into a form applicable to downstream tasks by calculating a relation matrix, constructing an association graph, and constructing a scope delimitation.
[0018] Preferably, it also includes building a large-scale computing engine based on distributed computing technologies Spark and Flink, which performs hierarchical processing of massive graph knowledge according to business needs, performs real-time, near real-time and offline computing on different knowledge, improves computing speed, and thus dynamically updates the large-scale graph in real time.
[0019] The beneficial effects of this invention include:
[0020] This invention integrates multiple data sources and various types of knowledge in the financial field, solving the problems of data dispersion, redundancy, and inconsistency in key financial knowledge repositories. It provides a more comprehensive and consistent knowledge view, enabling large-scale data sharing and interaction, and making financial knowledge querying and reasoning more efficient. Furthermore, the automated and semi-automated knowledge extraction and association processes in this invention (automated refers to automatically acquiring financial data using scripts, where the rules for data acquisition can be specified according to needs; semi-automated refers to the introduction of experts for secondary expansion and correction when the graph is expanded based on deep learning algorithms for entity extraction and association, and when the graph is expanded based on common sense knowledge) effectively reduce the workload of manual processing, improve the efficiency of graph construction, and allow the knowledge graph to grow with the increase of data and the expansion of fields through regular updates and maintenance, exhibiting strong scalability and adaptability. In addition, the construction of large-scale financial knowledge graphs has broad application prospects in fields such as user profiling and relationship mining, risk assessment, and supply chain finance. Attached Figure Description
[0021] Figure 1 This is an architecture diagram of a method for constructing a large-scale knowledge graph for the financial field, as shown in Example 1.
[0022] Figure 2 A detailed flowchart for constructing a large-scale knowledge graph in Example 1 is provided.
[0023] Figure 3 This is a roadmap of commonly used knowledge graph construction techniques in Example 1.
[0024] Figure 4 This is a technical roadmap for the common sense knowledge base structure in Example 1.
[0025] Figure 5 Example 1 illustrates a knowledge expansion method based on a common sense knowledge base.
[0026] Figure 6 This is a technology roadmap for the knowledge graph representation learning module in Example 1. Detailed Implementation
[0027] 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 a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0028] Example 1
[0029] The following is in conjunction with the appendix Figure 1-6 Specific embodiments of the present invention will be described in detail;
[0030] A method for constructing large-scale knowledge graphs in the financial sector, comprising:
[0031] Data Layer: The high-performance distributed data processing platform Spark integrates with the Quandl open API, a structured open knowledge base. [Quandl is an open platform for financial and economic data, providing financial and economic data from various data providers, including stocks, futures, indices, macroeconomic data, etc. Additionally, Quandl offers rich APIs and data download services.] 347,684,407 semi-structured data entries from Hong Kong and Shanghai / Shenzhen listed companies' trading data were collected, along with 1,738,730,800 structured data entries from Eastmoney and Sina Finance, including registration and approval data, fund flows, futures and options data, fund data, and annual / quarterly reports. Data persistence operations were performed on the Spark platform, and Spark SQL was used to load the data and perform data preprocessing tasks, including removing duplicate data, handling missing values, and mapping data to a common format. Feature engineering was then conducted to extract relevant features from the raw data. The streaming processing framework Apache Kafka was used to acquire and update streaming data, including tagged data, business data, and transaction data, in real time. Standards and fields for data acquisition were defined to facilitate data cleaning, transformation, and formatting, ensuring data quality and consistency.
[0032] Knowledge Layer: 1. Concept and knowledge extraction are performed on the data acquired from the data layer, followed by graph construction, knowledge fusion, and knowledge reasoning to form a preliminary large-scale graph platform. The data is then stored in the distributed graph database Neo4j, and knowledge bases and concept bases are formed according to different data types, providing preliminary data management and query services. The large-scale knowledge graph created by this invention contains 8.2 million entities and 19.5 million relations. Specifically, the detailed process of large-scale knowledge graph construction is as follows... Figure 2 As shown. Wherein:
[0033] Knowledge extraction includes entity recognition and relation extraction. It employs a long-range supervised convolutional deep neural network for knowledge extraction, solving the problems of automatic generation of training corpora and data noise in open domain environments. It also uses a deep convolutional neural network to solve the text semantic representation problem in the knowledge acquisition process, reducing the excessive reliance of traditional methods on natural language processing tools and the problem of error accumulation. The end-to-end knowledge base question answering and reasoning method based on reinforcement learning transforms the traditional symbol-based question answering and reasoning process into a learnable problem based on semantic numerical computation, thereby performing entity correlation calculation and knowledge reasoning.
[0034] 2. Construction of a Common Sense Knowledge Base: Common sense knowledge is an important component of large-scale knowledge graphs. Based on the construction of large-scale knowledge graphs, this invention constructs a common sense knowledge base, then expands the knowledge graph using rules from the common sense knowledge base. Next, graph representation technology is applied, combining the rules from the common sense knowledge base to represent the knowledge graph. Finally, the graph association information obtained from the graph representation and the target association information provided by the common sense knowledge base are fed into a deep neural network to assist in its training. The specific scheme is as follows... Figure 3 As shown.
[0035] To integrate the vast but scattered and isolated domain common sense knowledge, this embodiment establishes a common sense knowledge base centered on an instance library, a rule library, and an association library. The structure of the common sense knowledge base is designed as follows: Figure 4 As shown; among them, the construction of the instance library mainly includes the representation of common sense instances and the acquisition of common sense knowledge; the construction of the rule library mainly includes the construction of a financial domain lexical ontology and the formulation of rules; the construction of the association library mainly involves the calculation of the target similarity matrix.
[0036] 3. Knowledge Expansion of the Common Sense Knowledge Base: The knowledge expansion method based on the common sense knowledge base is entity-independent and does not rely on the vector representation of entities. Its logical rules can be applied to any newly added entity in the knowledge base, exhibiting strong universality. Therefore, to expand the knowledge graph of common sense knowledge, this embodiment employs two reasoning processes: forward reasoning and backward reasoning. Forward reasoning uses existing facts (relation triples) in the knowledge graph to match the preconditions of rules. When the preconditions of a rule are satisfied, the facts of its conclusion are added to the knowledge base for expansion. The backward reasoning process is the opposite of forward reasoning. First, a target is assumed. Then, a rule whose conclusion matches the target is found. Next, the preconditions of the rule are recursively proven to be true. Once proven true, the facts of its conclusion are added to the knowledge base for expansion.
[0037] 4. Common Sense Knowledge - Knowledge Graph Joint Representation Learning: The graph representation learning module for large-scale knowledge graphs is mainly divided into three layers: transformation layer, model layer, and interface layer. The transformation layer transforms the original knowledge graph into a standard graph and obtains its adjacency matrix representation, while providing a lookup table for easy querying of vertex representations. The model layer takes the adjacency matrix as input and uses the graph representation learning model to obtain a vector representation matrix composed of low-dimensional vector representations of each vertex. The interface layer, according to the actual task requirements, obtains the vector representation of the classification labels through the lookup table, transforms it into a probability distribution, and further transforms it into a graph association matrix as a bias for the classification results of the deep learning module, indicating the learning process and accelerating convergence. The overall module framework is as follows: Figure 6 As shown. The transformation layer and model layer are key modules in this module for implementing knowledge graph identifier learning, while the interface layer is a data interface created to facilitate downstream tasks. Specifically:
[0038] The transformation layer is responsible for preprocessing the knowledge graph, converting it into a standard graph for subsequent graph representation learning models. Given a knowledge graph... (H,R,T) , where H is the head entity set, R is the tail entity set, T is the relation set, and all predicates involved in the knowledge reasoning rules constitute the set U.
[0039] The model layer applies a graph representation learning model to the preprocessed standard graph. This embodiment uses the SDNE model, which has performed well in various practical tasks. It has been proven to have a powerful ability to learn complex data structures and has achieved substantial success in processing image, text, and audio data. The SDNE model employs a semi-supervised deep model for graph representation learning, using a deep architecture to capture highly nonlinear network topologies. It consists of multiple nonlinear mapping functions that capture network structures by mapping input data to a highly nonlinear latent space. On one hand, the model reconstructs the neighborhood structure of each vertex through unsupervised learning to preserve local characteristics. On the other hand, the model uses Laplacian mappings to learn global graph characteristics through supervised learning, using first-order similarity as supervisory information. These global characteristics primarily reflect semantic features.
[0040] Interface layer: By calculating the relationship matrix, constructing the association graph, and defining the scope, the graph representation is transformed into a form that can be applied to downstream tasks, providing services such as querying, retrieval, and matching. Specifically, it can be used in interactive human-computer dialogue systems, user panoramic profiles, and relationship mining.
[0041] Example 2
[0042] The difference between this embodiment and Embodiment 1 is that it includes large-scale dynamic real-time updates of the knowledge graph. Specifically, to address the problems of slow and inefficient real-time computation of large-scale knowledge graphs, a large-scale computing engine is built based on the distributed computing technologies Spark and Flink. The massive knowledge graph is processed in a hierarchical manner according to business needs, and different knowledge is computed in real-time, near real-time, and offline. This improves the computation speed of graph features and cognition, and keeps the knowledge query and retrieval speed at around 10ms.
[0043] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.
Claims
1. A method for constructing large-scale knowledge graphs in the financial field, characterized in that, include: S1: Obtain financial and economic data from open knowledge bases, perform concept extraction and knowledge extraction, and then construct a large-scale knowledge graph; S2: Build a commonsense knowledge base based on a large-scale knowledge graph; S3: Expand the knowledge base of common sense using large-scale knowledge graphs; S4: Utilize the graph representation learning module to perform joint representation learning of common sense knowledge base and large-scale knowledge graph; In S1, obtaining financial and economic data from an open knowledge base specifically involves using the streaming processing framework Apache Kafka to acquire and update streaming data, including tagged data, business data, and transaction data, in real time, and defining the standards and fields for data acquisition to facilitate data cleaning, transformation, and formatting. The knowledge extraction includes entity recognition and relation extraction, and the knowledge extraction process includes: We employ a deep convolutional neural network to address the text semantic representation problem in the knowledge acquisition process, thereby reducing the excessive reliance on natural language processing tools and the problem of error accumulation in traditional methods. An end-to-end knowledge base question answering and reasoning method based on reinforcement learning is adopted to transform the traditional question answering and reasoning process based on symbolic representation into a learnable problem based on semantic numerical computation, thereby performing entity correlation calculation and knowledge reasoning. The knowledge expansion in S3 includes forward reasoning and backward reasoning. Forward reasoning uses existing facts in the knowledge graph to match the premises of a rule. When the premises of a rule are met, the facts of its conclusion are added to the common sense knowledge base for expansion. The existing facts include relation triples. Backward reasoning first assumes a target and finds a rule whose conclusion matches the target. It recursively proves that the premises of the rule are all true. When the proof is true, the facts of its conclusion are added to the common sense knowledge base. The graph representation learning module in S4 includes a transformation layer, a model layer, and an interface layer. The transformation layer transforms the original knowledge graph into a standard graph and obtains its adjacency matrix representation, while providing a lookup table for easy querying of vertex representations. The model layer takes the adjacency matrix as input and obtains a vector representation matrix composed of low-dimensional vector representations of each vertex through the graph representation learning model. The interface layer obtains the vector representation of the classification label through the lookup table according to the actual task requirements, transforms it into a probability distribution, and further transforms it into a graph association matrix as a bias of the classification result of the deep learning module to indicate the learning process and accelerate convergence.
2. The method for constructing a large-scale knowledge graph for the financial field according to claim 1, characterized in that, The common sense knowledge base described in S2 includes an instance base, a rule base, and an association base. The instance base includes common sense instance representation and common sense knowledge acquisition; the rule base includes constructing a financial domain lexical ontology and formulating rules; and the association base includes calculating the target similarity matrix.
3. The method for constructing a large-scale knowledge graph for the financial field according to claim 1, characterized in that, The transformation layer is used to preprocess the knowledge graph into a standard graph to facilitate subsequent input into the graph representation learning model, including: given a knowledge graph (H,R,T) , where H is the head entity set, R is the tail entity set, T is the relation set, and all predicates involved in the knowledge reasoning rules constitute the set U.
4. The method for constructing a large-scale knowledge graph for the financial field according to claim 1, characterized in that, The model layer includes a graph representation learning model. The process of applying the graph representation learning model includes: first, reconstructing the neighborhood structure of each vertex to preserve local characteristics through unsupervised learning; second, using Laplace mapping to learn the global characteristics of the graph further by using first-order similarity as supervision information through supervised learning.
5. The method for constructing a large-scale knowledge graph for the financial field according to claim 1, characterized in that, The interface layer transforms the graph representation into a form applicable to downstream tasks by calculating the relationship matrix, constructing the association graph, and defining the scope.
6. The method for constructing a large-scale knowledge graph for the financial field according to claim 1, characterized in that, It also includes building a large-scale computing engine based on distributed computing technologies Spark and Flink, which performs hierarchical processing of massive graph knowledge according to business needs, performs real-time, near real-time and offline computing on different knowledge, improves computing speed, and thus dynamically updates the large-scale graph in real time.