Artificial intelligence big data semantic understanding and reasoning method and system fused with knowledge graph

Abstract

The application belongs to the field of artificial intelligence, in particular to an artificial intelligence big data semantic understanding and reasoning method and system fusing a knowledge graph, and the method comprises the following steps: S1, multi-source data fusion and preprocessing: collecting and fusing multi-source heterogeneous data, and performing cleaning, alignment and formatting; S2, dynamic causal knowledge graph construction: performing deep semantic analysis on the preprocessed data based on a pre-training language model, introducing a causal test algorithm to mine implicit causal relationships, and combining sentiment analysis to label entity attributes, to generate a dynamic knowledge graph containing entities, relationships, causal weights and sentiment labels; S3, multi-modal semantic feature and structural feature joint coding, by introducing causal test and sentiment analysis, the knowledge graph not only stores facts, but also contains causal logic and sentiment tendency between events, greatly improving the machine's ability to analyze complex semantics.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for semantic understanding and reasoning of big data in artificial intelligence that integrates knowledge graphs. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, accurately analyzing user intent and performing deep reasoning from massive, multi-source, and heterogeneous data has become a core challenge. Traditional semantic understanding methods mostly rely on keyword matching or shallow machine learning models, making it difficult to capture complex semantic relationships and implicit causal logic. In recent years, question-answering and reasoning systems integrating knowledge graphs have made some progress, but the following technical shortcomings still exist: Insufficient semantic understanding depth: Existing knowledge graph constructions mostly focus on extracting entities and explicit relationships, lacking the mining of deep semantics such as causal logic and sentiment between events, resulting in limited expressive power of the graph.

[0003] Fragmented reasoning mechanisms: Graph-based reasoning (such as path sequencing) and neural network-based semantic reasoning often proceed independently and are difficult to integrate and complement. When encountering knowledge gaps, the reasoning chain is prone to breakage and lacks robustness.

[0004] Lagging knowledge timeliness: Once a static knowledge graph is built, it is difficult to update dynamically and cannot adapt to changes in data flow, resulting in outdated reasoning results that cannot meet the needs of real-time decision-making. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies and propose a method and system for semantic understanding and reasoning based on artificial intelligence big data that integrates knowledge graphs.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The method for semantic understanding and reasoning of big data in artificial intelligence that integrates knowledge graphs includes the following steps: S1. Multi-source data fusion and preprocessing: Collect and fuse heterogeneous data from multiple sources, and perform cleaning, alignment and formatting; S2. Construction of dynamic causal knowledge graph: Based on the pre-trained language model, deep semantic analysis is performed on the pre-processed data. Causal testing algorithm is introduced to mine implicit causal relationships. Combined with sentiment analysis to label entity attributes, a dynamic knowledge graph containing entities, relationships, causal weights and sentiment tags is generated. S3. Joint encoding of multimodal semantic features and structural features: The structural features of the knowledge graph and the semantic features of the entities are encoded separately to obtain a joint feature representation; S4. Semantic-Structure Collaborative Reasoning: Construct an adaptive gated fusion reasoning engine, dynamically fuse structural vectors and semantic vectors for multi-hop reasoning, and trigger a large model gap filling mechanism when the reasoning link is broken to generate a complete reasoning path. S5. Reasoning Result Verification and Feedback Collection: Compare the reasoning results with actual data, calculate reward signals, and record knowledge gaps; S6. Knowledge graph self-evolution based on reinforcement learning: Construct a dual-mode driven knowledge evolution agent, dynamically decide on the update operation of the knowledge graph according to the reward signal, and feed the updated graph back to S2 to form a closed-loop evolution.

[0007] Preferably, in step S2, the Granger causality test algorithm combined with an attention mechanism is introduced to mine the implicit causal relationship between events, and the semantic sentiment analysis results are labeled as dynamic attributes of entities.

[0008] Preferably, the adaptive gating fusion inference engine in step S4 includes a structural feature encoding channel based on a graph neural network and a semantic feature encoding channel based on a knowledge-aware attention mechanism, which are dynamically fused through a gating unit; when the inference path is broken, a large model completion mechanism is triggered to generate relay nodes to form a complete inference chain.

[0009] Preferably, the knowledge evolution agent in step S6 adopts a reinforcement learning strategy, and dynamically determines the weight adjustment, addition or elimination operations of entities and relations in the knowledge graph by monitoring changes in external data sources and reward signals in step S5.

[0010] Preferably, the multi-source heterogeneous data in step S1 includes, but is not limited to, structured data, semi-structured data, and unstructured data; the cleaning includes missing value processing, outlier detection and correction; the alignment includes entity alignment and conflict resolution; and the formatting is converted into a unified graph data model.

[0011] Preferably, the joint encoding of multimodal semantic features and structural features in step S3 specifically involves: extracting structural features of the knowledge graph through a graph convolutional network or a graph attention network to obtain a structural vector; simultaneously, encoding the textual description of entities through a pre-trained language model to obtain a semantic vector; and fusing the structural vector and the semantic vector through concatenation or weighted summation to obtain a joint feature representation.

[0012] Preferably, the reward signal calculation process in step S5 includes: performing precise matching or fuzzy matching between the reasoning path or result and real-world event data, and assigning positive or negative rewards based on the matching degree; the knowledge gap record includes missing entities, relationships, or incorrect causal weights.

[0013] Preferably, the dual-mode driven knowledge evolution agent in step S6 includes an exploration mode and an exploitation mode: in the exploration mode, the agent attempts to add new entities or relationships based on changes in external data sources; in the exploitation mode, the agent strengthens or weakens the weights of existing entities and relationships based on reward signals; the update operation is executed atomically through a knowledge graph query language or graph database transaction.

[0014] Preferably, the Granger causality test algorithm combined with the attention mechanism specifically involves: using the attention mechanism to filter out highly correlated time series event pairs, narrowing the computational scope of the Granger causality test, and then performing the Granger causality test on the filtered event pairs to determine the causal direction, causal strength, and time lag between events. The large model gap completion mechanism specifically involves: when the inference link is broken, concatenating the context information at the break point, the query intent, and the semantic features of the entities before and after the break into a prompt word, inputting it into the large-scale language model, generating potential relay entities or relations, and inserting them as virtual nodes into the current inference path, which are then converted into permanent nodes after subsequent verification and confirmation.

[0015] This invention also proposes an artificial intelligence big data semantic understanding and reasoning system that integrates knowledge graphs, including: The data fusion module collects and merges heterogeneous data from multiple sources, and performs cleaning, alignment, and formatting. The knowledge graph construction module performs deep semantic analysis on pre-processed data based on a pre-trained language model, introduces a causal testing algorithm to mine implicit causal relationships, and combines sentiment analysis to label entity attributes, generating a dynamic knowledge graph containing entities, relationships, causal weights, and sentiment tags. The feature encoding module encodes the structural features of the knowledge graph and the semantic features of the entities respectively, to obtain a joint feature representation; The collaborative reasoning module constructs an adaptive gated fusion reasoning engine, dynamically fuses structural vectors and semantic vectors for multi-hop reasoning, and triggers a large model gap filling mechanism when the reasoning link is broken to generate a complete reasoning path. The feedback collection module compares the reasoning results with the actual data, calculates the reward signal, and records the knowledge gap; The knowledge evolution module constructs a dual-mode driven knowledge evolution agent, which dynamically decides on the update operation of the knowledge graph based on the reward signal, and feeds the updated graph back to S2 to form a closed-loop evolution.

[0016] The beneficial effects of the artificial intelligence big data semantic understanding and reasoning method and system that integrates knowledge graphs in this invention are as follows: Enhanced semantic understanding: By introducing causal testing and sentiment analysis, knowledge graphs not only store facts, but also contain causal logic and sentiment tendencies between events, greatly improving the machine's ability to parse complex semantics.

[0017] Improved robustness and coherence of reasoning: By synergistically integrating semantic and structural features and introducing a large model gap-filling mechanism, reasoning interruptions caused by data sparsity are effectively avoided, ensuring the integrity and reliability of reasoning logic.

[0018] Knowledge Dynamic Adaptation: The knowledge self-evolution mechanism based on reinforcement learning enables the knowledge graph to be dynamically updated as the data flow changes, and the inference results always remain timely, realizing the system's continuous learning and self-optimization. Attached Figure Description

[0019] Figure 1 This is a flowchart of the AI ​​big data semantic understanding and reasoning method that integrates knowledge graphs proposed in this invention. Figure 2 This is a block diagram of the AI ​​big data semantic understanding and reasoning system that integrates knowledge graphs, as proposed in this invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0021] Reference Figures 1-2 The method for semantic understanding and reasoning based on artificial intelligence big data, which integrates knowledge graphs, includes the following steps: S1. Multi-source data fusion and preprocessing: Collect and fuse heterogeneous data from multiple sources, and perform cleaning, alignment and formatting; S2. Construction of dynamic causal knowledge graph: Based on the pre-trained language model, deep semantic analysis is performed on the pre-processed data. Causal testing algorithm is introduced to mine implicit causal relationships. Combined with sentiment analysis to label entity attributes, a dynamic knowledge graph containing entities, relationships, causal weights and sentiment tags is generated. S3. Joint encoding of multimodal semantic features and structural features: The structural features of the knowledge graph and the semantic features of the entities are encoded separately to obtain a joint feature representation; S4. Semantic-Structure Collaborative Reasoning: Construct an adaptive gated fusion reasoning engine, dynamically fuse structural vectors and semantic vectors for multi-hop reasoning, and trigger a large model gap filling mechanism when the reasoning link is broken to generate a complete reasoning path. S5. Reasoning Result Verification and Feedback Collection: Compare the reasoning results with actual data, calculate reward signals, and record knowledge gaps; S6. Knowledge graph self-evolution based on reinforcement learning: Construct a dual-mode driven knowledge evolution agent, dynamically decide on the update operation of the knowledge graph according to the reward signal, and feed the updated graph back to S2 to form a closed-loop evolution.

[0022] In this embodiment, step S2 introduces the Granger causality test algorithm combined with an attention mechanism to mine the implicit causal relationship between events, and uses the semantic sentiment analysis results as dynamic attributes of entities for labeling.

[0023] In this embodiment, the adaptive gating fusion inference engine in step S4 includes a structural feature encoding channel based on a graph neural network and a semantic feature encoding channel based on a knowledge-aware attention mechanism, which are dynamically fused through a gating unit. When the inference path is broken, a large model completion mechanism is triggered to generate relay nodes and form a complete inference chain.

[0024] In this embodiment, the knowledge evolution agent in step S6 adopts a reinforcement learning strategy. By monitoring changes in external data sources and the reward signal in step S5, it dynamically determines the weight adjustment, addition, or elimination operations for entities and relations in the knowledge graph.

[0025] In this embodiment, the multi-source heterogeneous data in step S1 includes, but is not limited to, structured data, semi-structured data and unstructured data; cleaning includes missing value processing, outlier detection and correction; alignment includes entity alignment and conflict resolution; and formatting is converted into a unified graph data model.

[0026] In this embodiment, the joint encoding of multimodal semantic features and structural features in step S3 is specifically as follows: extracting structural features of the knowledge graph through a graph convolutional network or a graph attention network to obtain a structural vector; simultaneously, encoding the textual description of the entity through a pre-trained language model to obtain a semantic vector; and fusing the structural vector and the semantic vector by concatenation or weighted summation to obtain a joint feature representation.

[0027] In this embodiment, the reward signal calculation process in step S5 includes: performing precise matching or fuzzy matching between the reasoning path or result and real-world event data, and assigning positive or negative rewards based on the matching degree; the knowledge gap record includes missing entities, relationships, or erroneous causal weights.

[0028] In this embodiment, the dual-mode driven knowledge evolution agent in step S6 includes an exploration mode and an exploitation mode: in the exploration mode, the agent attempts to add new entities or relationships based on changes in external data sources; in the exploitation mode, the agent strengthens or weakens the weights of existing entities and relationships based on reward signals; the update operation is executed atomically through knowledge graph query language or graph database transactions.

[0029] In this embodiment, the Granger causality test algorithm combined with the attention mechanism is specifically as follows: the attention mechanism is used to filter out highly correlated time series event pairs, narrowing the calculation scope of the Granger causality test, and then the Granger causality test is performed on the filtered event pairs to determine the causal direction, causal strength and time lag between events. The large model gap completion mechanism is specifically as follows: when the inference link is broken, the context information at the break point, the query intent and the semantic features of the entities before and after the break are concatenated into a prompt word, which is input into the large-scale language model to generate potential relay entities or relations, and inserted as virtual nodes into the current inference path. After subsequent verification and confirmation, it is transformed into a permanent node.

[0030] This invention also proposes an artificial intelligence big data semantic understanding and reasoning system that integrates knowledge graphs, including: The data fusion module collects and merges heterogeneous data from multiple sources, and performs cleaning, alignment, and formatting. The knowledge graph construction module performs deep semantic analysis on pre-processed data based on a pre-trained language model, introduces a causal testing algorithm to mine implicit causal relationships, and combines sentiment analysis to label entity attributes, generating a dynamic knowledge graph containing entities, relationships, causal weights, and sentiment tags. The feature encoding module encodes the structural features of the knowledge graph and the semantic features of the entities respectively, to obtain a joint feature representation; The collaborative reasoning module constructs an adaptive gated fusion reasoning engine, dynamically fuses structural vectors and semantic vectors for multi-hop reasoning, and triggers a large model gap filling mechanism when the reasoning link is broken to generate a complete reasoning path. The feedback collection module compares the reasoning results with the actual data, calculates the reward signal, and records the knowledge gap; The knowledge evolution module constructs a dual-mode driven knowledge evolution agent, which dynamically decides on the update operation of the knowledge graph based on the reward signal, and feeds the updated graph back to S2 to form a closed-loop evolution. Example 2

[0031] The difference between this embodiment and Embodiment 1 is that: The method for semantic understanding and reasoning of big data in artificial intelligence that integrates knowledge graphs includes the following steps: S1. Multi-source data fusion and preprocessing: Collect and fuse multi-source heterogeneous data, use a sliding window mechanism to dynamically align real-time data streams, perform missing value processing, outlier detection and correction, entity alignment and conflict resolution, and convert it into a unified graph data model. S2. Construction of Probabilistic Graphical Causal Knowledge Graph: Based on variational autoencoder, deep semantic analysis is performed on the preprocessed data. Bayesian network structure learning algorithm is introduced and mutual information is combined to mine the probabilistic causal dependencies between entities. Sentiment analysis is used to label the sentiment attributes of entities, and a probabilistic graphical knowledge graph containing entities, relationships, causal probabilities and sentiment tags is generated. S3. Joint Encoding of Multimodal Semantic and Structural Features: Structural features of the knowledge graph are extracted through a graph attention network to obtain structural vectors; textual semantic features of entities are extracted through a bidirectional encoder (BERT) to obtain semantic vectors; and features are fused through a cross-modal attention mechanism to obtain joint feature representations. S4. Semantic-Structure Collaborative Reasoning: Construct a reasoning engine based on Monte Carlo tree search, dynamically simulate multiple reasoning paths, integrate structural vectors and semantic vectors to evaluate path value, and trigger a large model completion mechanism when the confidence of the reasoning path is low or the link is broken, generating candidate reasoning paths or relay nodes to form a complete reasoning chain. S5. Verification and Feedback Collection of Reasoning Results: Perform precise or fuzzy matching between the reasoning results and actual data, calculate reward signals based on the matching degree, and record knowledge gaps, including causal probability bias, missing entities or relationships. S6. Knowledge Graph Self-Evolution Based on Reinforcement Learning: Construct a knowledge evolution agent based on policy gradients, adopting an Actor-Critic architecture. It dynamically decides on the update operation of the knowledge graph according to the reward signal, including adjusting causal probabilities, adding or eliminating entities and relations, and feeding the updated graph back to S2 to form a closed-loop evolution. Example 3

[0032] The difference between this embodiment and Embodiment 1 is that: The method for semantic understanding and reasoning of big data in artificial intelligence that integrates knowledge graphs includes the following steps: S1. Multi-source financial data fusion and preprocessing: Collect and fuse structured data (including transaction records, financial statements, macroeconomic indicators), semi-structured data (including news announcements, company disclosures), and unstructured data (including research reports, public opinion texts, and social media comments) in the financial field, perform cleaning, entity alignment, timestamp standardization, and conflict resolution, and convert them into a unified graph data model; S2. Construction of Dynamic Temporal Causal Knowledge Graph: Based on the FinBERT pre-trained language model in the financial field, deep semantic analysis is performed on the pre-processed data. A vector autoregression model (VAR) is introduced in combination with Granger causality test to explore the temporal causal dependencies between financial events. The sentiment lexicon of finance (such as Loughran-McDonald sentiment lexicon) is used to label the sentiment of entities and generate a dynamic temporal knowledge graph containing entities, relations, causal lag order, influence intensity and sentiment labels. S3. Joint Encoding of Multimodal Semantic and Structural Features: Structural features of the knowledge graph are extracted through graph convolutional networks to obtain structural vectors; textual semantic features of entities are extracted through FinBERT to obtain semantic vectors; temporal features are extracted using temporal convolutional networks (TCN) to obtain temporal vectors; and structural vectors, semantic vectors, and temporal vectors are dynamically fused through a gating fusion mechanism to obtain a joint feature representation. S4. Semantic-Structure Collaborative Reasoning: Construct a hybrid reasoning engine based on Graph Attention Network (GAT) and Long Short-Term Memory Network (LSTM), dynamically integrate structural vectors, semantic vectors and temporal features to perform multi-hop reasoning, predict the development trend of financial events or risk propagation path, and trigger a large model gap filling mechanism when the reasoning link is broken to generate potential relay events or relationships and form a complete reasoning chain. S5. Validation and Feedback Collection of Reasoning Results: Compare the reasoning results (such as risk event prediction and correlation reasoning) with actual financial event data, calculate reward signals (including prediction accuracy, recall and causal lag matching degree), and record knowledge gaps, such as missing correlations, incorrect causal lag orders or sentiment label bias. S6. Knowledge Graph Self-Evolution Based on Reinforcement Learning: Construct a dual-mode driven knowledge evolution agent, which combines market dynamics and reward signals to dynamically decide on the weight adjustment, addition, or elimination of entities and relationships in the knowledge graph, and feeds the updated graph back to S2 to form a closed-loop evolution to adapt to changes in the financial market.

[0033] Test case Risk Transmission Reasoning and Prediction in the Financial Sector 1. Experimental Setup Experimental objective: To evaluate the semantic understanding accuracy, reasoning path completeness, and prediction accuracy of each method in the complex scenario of "negative public opinion about listed companies leading to risk transmission to upstream and downstream enterprises in the supply chain".

[0034] Dataset: Data sources: A-share listed company announcements, financial news, social media (stock forums, Weibo) public opinion, and financial statement data from 2018 to 2023.

[0035] Data scale: Includes 10,000 enterprise entities and 500,000 event / text data entries.

[0036] A set of standard answers was constructed: 2,000 risk transmission reasoning paths (e.g., "Supplier A suffers environmental penalties -> production is restricted -> supply to core customer B is delayed -> customer B defaults on product delivery -> customer B's stock price falls") were manually labeled by financial experts as a test set.

[0037] 2. Comparison of Model and Parameter Settings Baseline Model 1 (Traditional Method): Rule-based reasoning based on static knowledge graphs + keyword matching. It does not use dynamic updates or semantic analysis.

[0038] Baseline Model 2 (Deep Learning Method): Uses TransE for knowledge graph embedding and LSTM for time series prediction. However, it lacks external knowledge completion and feedback evolution mechanisms.

[0039] Example 1 (this patent): Dynamic causal graph + adaptive gating fusion + large model completion + reinforcement learning evolution.

[0040] Example 2 (this patent): Probabilistic causal graph + Monte Carlo tree search + large model completion + Actor-Critic evolution.

[0041] Example 3 (This patent - Financial Application): Financial Temporal Causal Graph + FinBERT + TCN Temporal Features + GAT-LSTM Hybrid Inference + Reinforcement Learning Evolution.

[0042] 3. Comparison of core indicators We selected three core indicators for quantitative comparison: Semantic Accuracy: The accuracy with which a model correctly understands the implicit semantic relationships between entities (such as "positive / negative", "causal / correlated").

[0043] Path Completeness: Given a head entity and a tail entity, the proportion of times the model can successfully complete the missing intermediate links (multi-hop reasoning).

[0044] Risk event prediction accuracy (Prediction Accuracy @7 days): The accuracy with which the model predicts related risk events that will occur within 7 days based on the current event.

[0045]

[0046] 4. Analysis of Experimental Results Comparison with baseline model: All three implementations significantly outperformed baseline models 1 and 2 across all metrics. In particular, the reasoning path completeness of implementation 1 improved by nearly 30 percentage points compared to baseline model 2, directly demonstrating the significant advantage of the large model gap completion mechanism in handling sparse knowledge graphs.

[0047] The improvement in semantic understanding accuracy is mainly due to the pre-trained language model and causal testing algorithm introduced in step S2, which can distinguish between "relevance" and "causality" and avoid logical errors in traditional methods.

[0048] Internal comparison between implementation examples: Example 3 performed best: achieving a risk prediction accuracy of 83.9%, significantly higher than Examples 1 and 2. This is attributed to: Domain-specific pre-training (FinBERT): Better understanding of financial terminology and context (e.g., "liquidity crunch" is more professional than "lack of money").

[0049] Temporal Feature Fusion (TCN): It captures the lag in risk transmission (such as the stock price falling 3 days after the penalty announcement), which was not explicitly modeled in Examples 1 and 2.

[0050] Dedicated sentiment dictionary: more accurately identifies the sentiment tendency of financial texts.

[0051] Example 2 is slightly better than Example 1: In terms of inference path completeness and prediction accuracy, Example 2 (Monte Carlo tree search + probabilistic graph) is slightly better than Example 1. This shows that the probabilistic graphical model is more robust than the deterministic dynamic graph when dealing with uncertain inference, and can generate multiple candidate paths for evaluation, rather than relying on only one optimal path.

[0052] The fusion framework proposed in Example 1 has significantly outperformed traditional methods in general scenarios. Its core architecture of "gated fusion + large model completion + reinforcement learning evolution" effectively solves the problems of semantic understanding depth and knowledge update lag.

[0053] Example 2 further enhances the robustness of the model's inference in uncertain environments by introducing probabilistic graphs and Monte Carlo tree search.

[0054] Example 3, which focuses on optimizations specific to the financial field (time series modeling, domain modeling), achieved the best results, verifying that the method of this patent has good domain adaptability and scalability.

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fusing knowledge graph and artificial intelligence big data semantic understanding and reasoning, characterized in that, Includes the following steps: S1. Multi-source data fusion and preprocessing: Collect and fuse heterogeneous data from multiple sources, and perform cleaning, alignment and formatting; S2. Construction of dynamic causal knowledge graph: Based on the pre-trained language model, deep semantic analysis is performed on the pre-processed data. Causal testing algorithm is introduced to mine implicit causal relationships. Combined with sentiment analysis to label entity attributes, a dynamic knowledge graph containing entities, relationships, causal weights and sentiment tags is generated. S3. Joint encoding of multimodal semantic features and structural features: The structural features of the knowledge graph and the semantic features of the entities are encoded separately to obtain a joint feature representation; S4. Semantic-Structure Collaborative Reasoning: Construct an adaptive gated fusion reasoning engine, dynamically fuse structural vectors and semantic vectors for multi-hop reasoning, and trigger a large model gap filling mechanism when the reasoning link is broken to generate a complete reasoning path. S5. Reasoning Result Verification and Feedback Collection: Compare the reasoning results with actual data, calculate reward signals, and record knowledge gaps; S6. Knowledge graph self-evolution based on reinforcement learning: Construct a dual-mode driven knowledge evolution agent, dynamically decide on the update operation of the knowledge graph according to the reward signal, and feed the updated graph back to S2 to form a closed-loop evolution.

2. The method for semantic understanding and reasoning of artificial intelligence big data based on knowledge graphs according to claim 1, characterized in that, In step S2, the Granger causality test algorithm is introduced in combination with an attention mechanism to explore the implicit causal relationship between events, and the semantic sentiment analysis results are used as dynamic attributes of entities for labeling.

3. The method for semantic understanding and reasoning of artificial intelligence big data based on knowledge graphs according to claim 2, characterized in that, The adaptive gating fusion inference engine in step S4 includes a structural feature encoding channel based on a graph neural network and a semantic feature encoding channel based on a knowledge-aware attention mechanism, which are dynamically fused through gating units. When the inference path is broken, a large model completion mechanism is triggered to generate relay nodes and form a complete inference chain.

4. The method for semantic understanding and reasoning of artificial intelligence big data based on knowledge graphs according to claim 3, characterized in that, The knowledge evolution agent in step S6 adopts a reinforcement learning strategy. By monitoring changes in external data sources and the reward signal in step S5, it dynamically determines the weight adjustment, addition, or elimination operations for entities and relations in the knowledge graph.

5. The artificial intelligence big data semantic understanding and reasoning method integrating knowledge graphs according to claim 4, characterized in that, The multi-source heterogeneous data in step S1 includes, but is not limited to, structured data, semi-structured data, and unstructured data; the cleaning includes missing value processing, outlier detection and correction; the alignment includes entity alignment and conflict resolution. The formatting is converted into a unified graph data model.

6. The artificial intelligence big data semantic understanding and reasoning method integrating knowledge graphs according to claim 5, characterized in that, The joint encoding of multimodal semantic features and structural features in step S3 specifically involves: extracting structural features of the knowledge graph through a graph convolutional network or a graph attention network to obtain a structural vector; simultaneously, encoding the textual description of entities through a pre-trained language model to obtain a semantic vector; and fusing the structural vector and the semantic vector through concatenation or weighted summation to obtain a joint feature representation.

7. The artificial intelligence big data semantic understanding and reasoning method integrating knowledge graphs according to claim 6, characterized in that, The reward signal calculation process in step S5 includes: performing precise or fuzzy matching between the reasoning path or result and real-world event data, and assigning positive or negative rewards based on the matching degree; the knowledge gap record includes missing entities, relationships, or incorrect causal weights.

8. The method for semantic understanding and reasoning of artificial intelligence big data based on knowledge graphs according to claim 7, characterized in that, In step S6, the dual-mode driven knowledge evolution agent includes an exploration mode and an exploitation mode: in the exploration mode, the agent attempts to add new entities or relationships based on changes in external data sources; in the exploitation mode, the agent strengthens or weakens the weights of existing entities and relationships based on reward signals; the update operation is executed atomically through a knowledge graph query language or graph database transaction.

9. The artificial intelligence big data semantic understanding and reasoning method integrating knowledge graphs according to claim 8, characterized in that, The Granger causality test algorithm combined with the attention mechanism specifically works as follows: the attention mechanism is used to filter out highly correlated time series event pairs, narrowing the computational scope of the Granger causality test. Then, the Granger causality test is performed on the filtered event pairs to determine the causal direction, causal strength, and time lag between events. The large model gap completion mechanism specifically works as follows: when the inference link is broken, the context information at the break point, the query intent, and the semantic features of the entities before and after the break are concatenated into a prompt word, which is input into the large-scale language model to generate potential relay entities or relations. These are then inserted as virtual nodes into the current inference path and converted into permanent nodes after subsequent verification and confirmation.

10. An AI big data semantic understanding and reasoning system integrating knowledge graphs, characterized in that: include: The data fusion module is used to perform step S1 as described in claim 1; The map construction module is used to perform step S2 as described in claim 1; The feature encoding module is used to perform step S3 as described in claim 1; A collaborative reasoning module is used to execute step S4 as described in claim 1; A feedback collection module is used to perform step S5 as described in claim 1; The knowledge evolution module is used to perform step S6 as described in claim 1.

Images