A method for event information enhancement and deduction prediction based on knowledge graph

Abstract

The present invention relates to a method for event information enhancement and deduction prediction combined with knowledge graphs, and belongs to the fields of natural language processing and knowledge graphs. In order to solve the problem that the existing event deduction methods fail to effectively integrate multi-dimensional information, resulting in insufficient accuracy and robustness of the deduction results; traditional event deduction methods mostly rely on a large amount of entity link information, resulting in low efficiency in practical applications and susceptibility to data scarcity, the method of the present invention includes five steps: data input and preprocessing, knowledge graph construction and optimization, event semantic background enhancement, event deduction prediction, and event deduction prediction. The present invention can effectively overcome the limitations of traditional methods in event deduction by introducing technologies such as cross-domain knowledge graphs and graph neural networks. The present invention uses knowledge graphs to construct rich event semantic backgrounds, and combines the temporal and causal relationships between events to enhance and deduce event information, thereby improving the accuracy and reliability of event deduction.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing and knowledge graphs, and specifically relates to an event information enhancement and deduction prediction method combined with knowledge graphs. Background Art

[0002] Event deduction and prediction involves inferring future events based on historical data and related information. The goal is to accurately support decision-making by deducing current and future events based on an understanding and analysis of historical events and their interrelationships. This type of task is widely used in fields such as financial market analysis and legal case prediction. During event deduction, the system must combine multiple pieces of information, including temporal, causal, and spatial information, to infer and predict the subsequent development of events.

[0003] Existing event deduction techniques primarily rely on data-driven approaches, using machine learning or deep learning models to learn patterns from historical data and predict future events. However, traditional event deduction methods are typically based on static data and lack the ability to effectively integrate the deep semantic relationships and cross-domain knowledge behind events, resulting in insufficient performance when faced with complex and dynamic event deduction tasks. For example, in financial market forecasting, while machine learning models can extract certain trends from historical data, they lack an understanding of the causal relationships behind market volatility, making it difficult for the models to accurately predict market events. Furthermore, existing event deduction systems often rely on a single data source when processing event information, lacking the ability to comprehensively process multi-source and multi-dimensional information, making it difficult to fully reflect the temporal and spatial changes and causal relationships of events. Therefore, how to integrate cross-domain knowledge graphs with event deduction tasks to enhance event information deduction capabilities has become a major challenge in current technology.

[0004] This paper proposes a method for event information enhancement and prediction based on knowledge graphs. By incorporating cross-domain knowledge graphs and graph neural networks (GNNs), it effectively overcomes the limitations of traditional methods for event prediction. The innovation of this method lies in leveraging knowledge graphs to construct a rich semantic context for events and integrating the temporal and causal relationships between events to enhance and predict event information, thereby improving the accuracy and reliability of event prediction. Summary of the Invention

[0005] (1) Technical issues to be resolved

[0006] The technical problem to be solved by the present invention is how to provide an event information enhancement and deduction prediction method combined with knowledge graph, so as to solve the problem that existing event deduction methods usually fail to effectively integrate multi-dimensional information, resulting in insufficient accuracy and robustness of deduction results; traditional event deduction methods mostly rely on a large amount of entity link information, resulting in low efficiency of the system in practical applications and susceptibility to data scarcity.

[0007] (2) Technical solution

[0008] In order to solve the above technical problems, the present invention proposes a method for event information enhancement and deduction prediction combined with knowledge graph, characterized in that the method comprises the following steps:

[0009] Step 1: Data input and preprocessing: Extract event information from multi-source heterogeneous data, complete data cleaning and standardization, and unify the time and geographic data formats;

[0010] Step 2: Knowledge graph construction and optimization: Build a knowledge graph based on preprocessed data, including event nodes and their temporal, spatial, and causal relationship chains, and optimize the knowledge graph.

[0011] Step 3: Enhance the semantic context of events: Based on the knowledge graph, we use multi-dimensional semantic enhancement technology to generate temporal features through sine-cosine encoding, embed geographic locations using GeoEncoder, and extract potential causal chains based on a causal analysis model based on RoBERTa. Subsequently, we use a graph convolutional neural network (GCN) for multi-hop semantic propagation, and enhance the information aggregation ability of key nodes through the self-attention mechanism, thereby enhancing the global consistency of event representation.

[0012] Step 4: Event deduction and prediction: Based on the semantically enhanced knowledge graph, we use the graph convolutional neural network (GCN) to capture explicit node relationships and combine it with the graph attention network (GAT) to assign node weights to build an event deduction model. When deducing paths, the model uses the heuristic A* algorithm to find the optimal path and normalizes the path weights using the Softmax function to generate a probability distribution for future events. Then, we train the event deduction model.

[0013] Step 5: Dynamic Optimization and Result Output: The knowledge graph is updated in real time through consistency detection rules, quickly verifying and adding new nodes and relationships. Variational Bayesian inference is used to quantitatively model the uncertainty of events and provide confidence scores for the deduction results. In the output phase, prediction results are provided in JSON format, and the deduction path and knowledge graph structure are visualized using D3.js. A detailed event report is also generated to support real-world application scenarios.

[0014] Among them, in the step one, extracting event information from multi-source heterogeneous data includes: when extracting text features, using NER to identify event subjects, objects and actions, and combining BERT to encode time and causal clues; and when extracting image features, using YOLOv8 to detect key objects, and combining the I3D model for video understanding and action recognition to identify dynamic events, providing high-quality input for subsequent knowledge graph construction.

[0015] (3) Beneficial effects

[0016] The present invention proposes a method for event information enhancement and deduction prediction combined with knowledge graphs. The present invention discloses a method for event information enhancement and deduction prediction combined with knowledge graphs, which has the following advantages:

[0017] (1) Multi-dimensional information fusion improves the accuracy of event deduction: This invention significantly improves the event deduction model's ability to understand multiple information such as time, space, and cause and effect by constructing a cross-domain knowledge graph and introducing a multi-dimensional event information enhancement algorithm. This enables the system to deduce more accurately in complex event scenarios, thereby improving the accuracy and robustness of the deduction.

[0018] (2) Reduced reliance on entity linking information: This paper proposes an event reasoning technology that does not rely on a large amount of entity linking information. By combining graph neural networks with knowledge graphs, it effectively reduces the negative impact of entity linking errors on event reasoning results. This technology reduces reliance on annotated data and improves the applicability and stability of the model in practical applications, especially in scenarios where data is scarce or dynamically changing.

[0019] (3) Graph Neural Network Optimization: This paper uses graph neural networks to model event information, which can deeply explore the complex connections and evolution patterns between events, improving the reasoning ability and prediction accuracy during event deduction. When dealing with complex event relationships and long-term event predictions, the system demonstrates higher adaptability and stronger deduction accuracy, and can cope with dynamically changing event environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall framework of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings and examples.

[0022] The present invention proposes a method for event information enhancement and deduction prediction combined with a knowledge graph, which mainly solves the following two problems: First, existing event deduction methods usually fail to effectively integrate multi-dimensional information (such as time, space, causal relationships, etc.), resulting in insufficient accuracy and robustness of the deduction results. To this end, the present invention proposes an event information enhancement method based on a knowledge graph, which enhances the semantic background of event information through the fusion of multi-dimensional information, thereby improving the effect of event deduction; second, traditional event deduction methods often rely on a large amount of entity link information, resulting in low efficiency of the system in practical applications and susceptibility to data scarcity. The present invention proposes a deduction technology that does not require a large number of entity links. By introducing the optimized structure of graph neural networks and knowledge graphs, it reduces dependence on entity links, while improving the accuracy and training efficiency of event deduction models.

[0023] The present invention provides an event information enhancement and deduction prediction method combined with knowledge graph. The overall framework is as follows: Figure 1 As shown, it mainly includes five steps: data input and preprocessing, knowledge graph construction and optimization, event semantic background enhancement, event deduction and prediction, and event deduction and prediction:

[0024] Step 1: Data input and preprocessing: Extract event information from multi-source heterogeneous data, complete data cleaning and standardization, including denoising, filling missing values, and unifying time and geographic data formats (ISO 8601 and WGS84).

[0025] Among them, extracting event information from multi-source heterogeneous data includes: when extracting text features, using NER to identify event subjects, objects and actions, and combining BERT to encode time and causal clues; when extracting image features, using YOLOv8 to detect key objects, and combining the I3D model for video understanding and action recognition to identify dynamic events, providing high-quality input for subsequent knowledge graph construction.

[0026] Step 2: Knowledge graph construction and optimization: Build a knowledge graph based on preprocessed data, including event nodes and their time, space, and causal relationship chains, and optimize the knowledge graph.

[0027] Temporal relationships are generated through timestamp sorting, spatial relationships are calculated based on Euclidean distance, and causal relationships are identified through Granger causality tests. In the process of optimizing the constructed knowledge graph, Node2Vec is used to generate node vector representations, node semantics are learned through random walks, and a node degree-based pruning algorithm is combined to reduce redundancy and improve the efficiency and inference capabilities of the knowledge graph.

[0028] Step 3: Enhance the semantic context of events: Based on the knowledge graph, we employ multi-dimensional semantic enhancement technology, generating temporal features through sine-cosine encoding, embedding geographic locations using GeoEncoder, and extracting potential causal chains using a causal analysis model based on RoBERTa. Subsequently, we employ GCN for multi-hop semantic propagation, using a self-attention mechanism to enhance information aggregation capabilities at key nodes, thereby enhancing the global consistency of event representation.

[0029] Step 4: Event Deduction and Prediction: Based on the semantically enhanced knowledge graph, a graph convolutional neural network (GCN) is used to capture explicit node relationships. Combined with a graph attention network (GAT), node weights are assigned to construct an event deduction model. When deducing paths, the model uses the heuristic A* algorithm to find the optimal path. Path weights are normalized using the Softmax function to generate a probability distribution for future events. The model is then trained, using a cross-entropy loss function to optimize predictions and improve deduction accuracy and robustness.

[0030] Step 5: Dynamic Optimization and Result Output: Consistency check rules are used to update the knowledge graph in real time, rapidly verifying and adding new nodes and relationships. Variational Bayesian inference is used to quantitatively model event uncertainty and provide confidence scores for prediction results. During the output phase, the system delivers prediction results in JSON format, visualizing the prediction path and knowledge graph structure using D3.js, and generating detailed event reports to support real-world applications. Example

[0031] The following is a detailed description of the five steps:

[0032] Step 1: Data input and preprocessing: The system extracts event-related information from multi-source heterogeneous data to build the basis for subsequent knowledge graphs.

[0033] During the data cleaning phase, data cleaning rules were applied, including removing redundant items, filling missing values, and correcting data format inconsistencies. Time data was standardized to the ISO 8601 standard format, and geographic location information was expressed using the WGS84 coordinate system.

[0034] For event feature extraction, a named entity recognition (NER) model is used to identify the subject, object, and action relationships of an event. Furthermore, a pre-trained BERT model is used to encode contextual features from event text descriptions, extracting temporal and causal clues. For image and video data, the YOLOv8 object detection model is used to extract key object features, which are then combined with the I3D action recognition algorithm to obtain dynamic event features. This processed data provides high-quality standardized input for subsequent knowledge graph construction.

[0035] Step 2: Knowledge Graph Construction and Optimization: After data preprocessing, the system constructs a knowledge graph based on the preprocessed data and optimizes its structure. Knowledge graph nodes include the subject and object of an event, as well as their associated attributes (time, location, causality, etc.). The relationship types between nodes are determined by predefined sets of temporal, spatial, and causal relationships. Temporal relationships construct temporal relationship chains by sorting event timestamps; spatial relationships calculate the Euclidean distance between nodes based on geographic coordinates, setting a threshold for spatial association; and causal relationships employ the Granger causality test for event pairs to identify triggers and causal chains between events. During graph optimization, the Node2Vec embedding algorithm is used to generate vector representations of nodes. Node semantic representations are learned through random walks and the Skip-Gram model, thereby enhancing the computational efficiency and retrieval capabilities of the knowledge graph. For sparse optimization, a node-degree-based pruning algorithm is used to remove low-degree nodes and their edges to reduce redundancy and improve inference efficiency.

[0036] Step 3: Enhance the semantic context of events: Based on the knowledge graph, the system uses a multi-dimensional semantic enhancement method to enhance the contextual semantic representation of event nodes. In the temporal dimension, periodic encoding techniques (such as sine-cosine encoding) are used to characterize timestamps and generate temporal context features for events. In the spatial dimension, GeoEncoder is used to embed geographic location information. In the causal dimension, a pre-trained causal relationship extraction model (such as a causal analysis model based on RoBERTa) is used to identify potential causal chains and add them to node attributes. Subsequently, a graph convolutional network (GCN) is used to propagate semantics across the knowledge graph, capturing implicit semantic connections between distant nodes through multi-layer message passing. Furthermore, the graph neural network structure employs a self-attention mechanism to enhance the information aggregation capabilities of high-importance nodes, thereby enhancing the global semantic consistency of events.

[0037] Step 4: Event Deduction and Prediction: Based on the semantically enhanced knowledge graph, the system uses an event deduction model to predict the likelihood and path of future events. The deduction model is based on a multi-layer graph neural network architecture. It first captures explicit relationships between nodes using GCN, and then uses a graph attention network (GAT) to assign weights to different nodes to highlight key event nodes. The model's deduction path search uses a heuristic A* algorithm, which optimizes the shortest path and combines relationship weights to find the most likely deduction path between events. To calculate the probability of future events, the model normalizes the cumulative weight of each path using a softmax function to generate a predicted distribution. The cross-entropy loss function is used during model training to compare the actual event chain labels with the predicted results for optimization, improving the accuracy and robustness of the deduction.

[0038] Step 5. Dynamic optimization and result output: In order to adapt to the update of dynamic events, the system has designed a real-time knowledge graph update mechanism. When new event data enters, the system automatically verifies the relationship between the new node and the existing graph through consistency detection rules, and quickly updates the graph structure through the newly added node module. At the same time, based on the variational Bayesian inference method, the uncertainty in the event data is quantitatively modeled to provide a confidence score for the deduction results, so that users can understand the uncertainty of the prediction. In terms of result output, the predicted events and their possible paths are output in JSON format, and a D3.js-based visualization display tool is provided to support interactive visualization of the knowledge graph structure and deduction path. In addition, the system generates a detailed event report, including the probability of the event, key influencing factors and deduction basis, which is convenient for application in actual scenarios.

[0039] This invention discloses a method for event information enhancement and deduction prediction combined with knowledge graph, which has the following advantages:

[0040] (1) Multi-dimensional information fusion improves the accuracy of event deduction: This invention significantly improves the event deduction model's ability to understand multiple information such as time, space, and cause and effect by constructing a cross-domain knowledge graph and introducing a multi-dimensional event information enhancement algorithm. This enables the system to deduce more accurately in complex event scenarios, thereby improving the accuracy and robustness of the deduction.

[0041] (2) Reduced reliance on entity linking information: This paper proposes an event reasoning technology that does not rely on a large amount of entity linking information. By combining graph neural networks with knowledge graphs, it effectively reduces the negative impact of entity linking errors on event reasoning results. This technology reduces reliance on annotated data and improves the applicability and stability of the model in practical applications, especially in scenarios where data is scarce or dynamically changing.

[0042] (3) Graph Neural Network Optimization: This paper uses graph neural networks to model event information, which can deeply explore the complex connections and evolution patterns between events, improving the reasoning ability and prediction accuracy during event deduction. When dealing with complex event relationships and long-term event predictions, the system demonstrates higher adaptability and stronger deduction accuracy, and can cope with dynamically changing event environments.

[0043] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for event information enhancement and deduction prediction combined with knowledge graph, characterized in that: The method comprises the following steps: Step 1: Data input and preprocessing: Extract event information from multi-source heterogeneous data, complete data cleaning and standardization, and unify the time and geographic data formats; Step 2: Knowledge graph construction and optimization: Build a knowledge graph based on preprocessed data, including event nodes and their temporal, spatial, and causal relationship chains, and optimize the knowledge graph. Step 3: Enhance the semantic context of events: Based on the knowledge graph, we use multi-dimensional semantic enhancement technology to generate temporal features through sine-cosine encoding, embed geographic locations using GeoEncoder, and extract potential causal chains based on a causal analysis model based on RoBERTa. Subsequently, we use a graph convolutional neural network (GCN) for multi-hop semantic propagation, and enhance the information aggregation ability of key nodes through the self-attention mechanism, thereby enhancing the global consistency of event representation. Step 4: Event deduction and prediction: Based on the semantically enhanced knowledge graph, we use the graph convolutional neural network (GCN) to capture explicit node relationships and combine it with the graph attention network (GAT) to assign node weights to build an event deduction model. When deducing paths, the model uses the heuristic A* algorithm to find the optimal path and normalizes the path weights using the Softmax function to generate a probability distribution for future events. Then, we train the event deduction model. Step 5: Dynamic optimization and result output: Update the knowledge graph in real time through consistency detection rules, and quickly verify and add new nodes and relationships; Use variational Bayesian inference to quantitatively model the uncertainty of events and provide confidence scores for the inference results; In the output phase, prediction results are provided in JSON format, and the deduction path and knowledge graph structure are visualized using D3.js. Detailed event reports are also generated to support real-world application scenarios. Among them, in the step one, extracting event information from multi-source heterogeneous data includes: when extracting text features, using NER to identify event subjects, objects and actions, and combining BERT to encode time and causal clues; and when extracting image features, using YOLOv8 to detect key objects, and combining the I3D model for video understanding and action recognition to identify dynamic events, providing high-quality input for subsequent knowledge graph construction.

2. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 1 is characterized in that: In step 1, the data cleaning stage adopts data cleaning rules, including removing redundant items, filling missing values ​​and correcting inconsistent data formats; the time data format is unified into the ISO 8601 standard format, and the geographic location information is represented by the WGS84 coordinate system.

3. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 1 is characterized in that: In the second step, the relationship type between nodes is determined by a predefined set of time relationship, spatial relationship and causal relationship; the time relationship is constructed by sorting the event timestamps to construct a time series relationship chain; Spatial relationships are based on calculating the Euclidean distance between nodes based on geographic coordinates and setting the threshold of spatial association; causal relationships use the Granger causality test method of event pairs to identify the triggering and causal chain between events.

4. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 3 is characterized in that: During the graph optimization process, the Node2Vec embedding algorithm is used to generate vector representations of nodes, and the node semantic representations are learned through random walks and Skip-Gram models to enhance the computational efficiency and retrieval capabilities of the knowledge graph.

5. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 3 is characterized in that: For sparse optimization, a node degree-based pruning algorithm is used to remove low-degree nodes and their edges to reduce redundancy.

6. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 1 is characterized in that: In step 3, the time dimension uses periodic sine-cosine coding technology to characterize the timestamp to generate the time context feature of the event; In the spatial dimension, GeoEncoder is used to embed geographic location information. In the causal dimension, a causal analysis model based on RoBERTa is used to identify potential causal chains and add them to node attributes. Subsequently, a graph convolutional neural network (GCN) is used to perform semantic propagation on the knowledge graph, capturing implicit semantic associations between distant nodes through multi-layer message passing.

7. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 1 is characterized in that: In step 4, the search for the deduction path adopts the heuristic A* algorithm, which takes the shortest path as the optimization goal and combines the relationship weights to find the most likely deduction path between events; in order to calculate the probability of future events, the cumulative weight of each path is normalized by the Softmax function to generate a predicted distribution.

8. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 7 is characterized in that: In step 4, the prediction results are optimized by the cross entropy loss function during the training phase to improve the accuracy and robustness of the deduction.

9. The event information enhancement and deduction prediction method combined with knowledge graph according to claim 1 is characterized in that: In step five, the predicted event and its possible path are output in JSON format, and a D3.js-based visualization tool is provided to support interactive visualization of the knowledge graph structure and deduction path. In addition, the system generates a detailed event report, including the probability of the event, key influencing factors, and deduction basis, to facilitate application in actual scenarios.

Images