Event causal mining method and system based on atlas constraint and timing optimal transmission

Abstract

The application provides an event causal relationship mining method and system based on atlas constraint and time sequence optimal transmission, which comprises the following steps: cutting a burst event description text into text blocks; performing atomic event extraction on each text block, and formalizing each atomic event into a binary feature group; mapping any two atomic events and into event nodes and in a preset knowledge atlas, calculating the shortest path topological distance of the event nodes and in the atlas, and constructing a teacher model probability distribution according to the shortest path topological distance; calculating the cosine distance between the vectors and as a basic semantic transmission cost, and performing time sequence constraint on the basic semantic transmission cost to construct a time sequence transmission cost matrix, searching for an optimal transmission matrix that minimizes the total transmission cost; and obtaining an event causal relationship that meets the time sequence optimal transmission under the constraint of a field knowledge atlas by minimizing a joint loss function of the teacher probability distribution and the optimal transmission matrix.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to an event causal mining method and system based on graph constraints and time-series optimal transmission. Background Technology

[0002] Emergency response and decision-making in the event of a sudden incident rely heavily on a deep understanding of the event's evolution. The related textual data is often unstructured or semi-structured, encompassing complex causal logic. However, the diverse data formats and scattered information make manual parsing and causal extraction costly, inefficient, and prone to human error. With the rapid development of artificial intelligence and natural language processing technologies, deep learning-based techniques have demonstrated powerful capabilities in text understanding and information extraction, providing new approaches to the intelligent parsing of unstructured data. Nevertheless, how to extract deep causal relationships from complex text, especially how to ensure the temporal order of causal inferences and their consistency with physical common sense at the algorithmic level, still faces many challenges.

[0003] Existing text-based knowledge unit association mining techniques mostly screen candidate knowledge unit pairs by calculating term frequency and distance features of knowledge units, and then use statistical indicators such as support to perform association mining. This method achieves automated association of textual knowledge to a certain extent, providing a statistical basis for the preliminary analysis of massive amounts of text; however, it is only suitable for shallow association analysis in general domains.

[0004] Existing technologies also propose domain-specific information verification techniques, which verify the authenticity of text content by constructing domain-specific knowledge graphs and feature engineering, but these techniques are only applicable to the detection of text in specific vertical categories.

[0005] With the development of large-scale modeling technology, existing technologies have proposed event extraction techniques based on large-scale models. These techniques leverage the semantic understanding capabilities of large language models to identify event trigger words and arguments from unstructured text, transforming them into structured data and reducing reliance on traditional rule templates. Although large-scale models improve the semantic generalization ability of event extraction, they have serious limitations when handling causal mining tasks. Specifically, general language models often lack strict physical constraints on the time dimension and are prone to making connections based solely on semantic similarity, thus producing reverse temporal logic errors such as "result leads to cause."

[0006] In summary, existing technologies for causal mining of sudden events generally suffer from a lack of hard constraints on physical time sequence and a lack of effective guidance from common-sense logic. Most existing statistical mining or semantic matching algorithms are time-insensitive, making it difficult to eradicate backward-chronological logical fallacies from the algorithm's underlying layers. Furthermore, existing methods fail to effectively utilize domain knowledge graphs as prior distributions to calibrate deep learning models, resulting in insufficient reasoning ability and logical interpretability when facing implicit causality or small-sample scenarios. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed to provide an event causal mining method and system based on graph constraints and time-optimal transmission to overcome the above problems.

[0008] This invention provides an event causality mining method based on graph constraints and time-series optimal transmission, the method comprising:

[0009] The text describing the emergency to be processed is divided into text blocks of a preset length, and overlapping areas are preserved between adjacent text blocks;

[0010] Atomic events are extracted from each text block to form atomic event groups, and each atomic event is formalized into a binary feature group. , Semantic content feature vectors used to characterize atomic events Timestamps used to characterize the timing of atomic events;

[0011] Any two atomic events in the atomic event group and Mapped to corresponding event nodes in a pre-defined domain knowledge graph and And calculate event nodes and In the domain knowledge graph, the shortest path topological distance is used to construct a teacher model probability distribution based on the shortest path topological distance between any two atomic events. This teacher model probability distribution is used to indicate the probability of events based on human common sense. Caused the incident The prior probability;

[0012] Calculate vectors and The cosine distance between them and treat it as an event To the event The basic semantic transmission cost is obtained by applying time constraints to the basic semantic transmission cost. The time transmission cost matrix is ​​constructed based on the time transmission costs of any two atomic events in the atomic event group. The optimal transmission matrix that minimizes the total transmission cost is searched using the Sinkhorn iterative algorithm to obtain the student model prediction probability distribution.

[0013] We construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model. By minimizing the joint loss function, we obtain the event causal relationship that satisfies the temporal optimal transmission under the domain knowledge graph constraint.

[0014] Another aspect of the present invention provides an event causality mining system based on graph constraints and time-series optimal transmission, the system comprising:

[0015] The preprocessing module is used to divide the text describing the emergency into text blocks of a preset length, while preserving overlapping areas between adjacent text blocks;

[0016] The event extraction module is used to extract atomic events from each text block to form atomic event groups, and to formalize each atomic event into a binary feature group. , Semantic content feature vectors used to characterize atomic events Timestamps used to characterize the timing of atomic events;

[0017] The common sense graph causal mining module is used to extract causal information from any two atomic events in an atomic event group. and Mapped to corresponding event nodes in a pre-defined domain knowledge graph and And calculate event nodes and In the domain knowledge graph, the shortest path topological distance is used to construct a teacher model probability distribution based on the shortest path topological distance between any two atomic events. This teacher model probability distribution is used to indicate the probability of events based on human common sense. Caused the incident The prior probability;

[0018] The temporal causal mining module is used to calculate vectors. and The cosine distance between them and treat it as an event To the event The basic semantic transmission cost is obtained by applying time constraints to the basic semantic transmission cost. The time transmission cost matrix is ​​constructed based on the time transmission costs of any two atomic events in the atomic event group. The optimal transmission matrix that minimizes the total transmission cost is searched using the Sinkhorn iterative algorithm to obtain the student model prediction probability distribution.

[0019] The joint optimization module is used to construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model. By minimizing the joint loss function, the event causal relationship that satisfies the temporal optimal transmission under the constraints of the domain knowledge graph is obtained.

[0020] This invention provides an event causal mining method and system based on graph constraints and temporal optimal transmission. By introducing a teacher probability distribution construction mechanism based on domain knowledge graphs, it creatively transforms discrete common-sense logic into continuous probability distributions, using these as "soft labels" to guide student model training. This allows the model to no longer rely solely on word co-occurrence statistics when faced with complex text lacking explicit connectives such as "cause" or "because," but instead to complete the logic using prior physical common sense. Compared to traditional statistical association mining methods, this invention effectively corrects the logical illusions easily generated by purely data-driven models, significantly improving the recall and accuracy of implicit causal relationships. Furthermore, this invention solves the temporal logic paradox problem in causal inference from the algorithmic level. Unlike existing standard optimal transmission algorithms that only consider semantic space distance while ignoring physical time constraints, this invention innovatively introduces a temporal causal masking mechanism into the transmission cost matrix. This mechanism can forcibly block all reverse-temporal probability transmission paths, ensuring that the mined causal chains logically and naturally satisfy the "causal preorder" axiom. This guarantees the physical consistency of the output results without the need for subsequent complex rule filtering, greatly improving the credibility of the mining results.

[0021] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0022] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings:

[0023] Figure 1 This is a flowchart of an event causal mining method based on graph constraints and time-series optimal transmission according to an embodiment of the present invention;

[0024] Figure 2 This is a structural block diagram of an event causal mining system based on graph constraints and time-series optimal transmission, according to an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined.

[0027] This invention proposes an event causal mining method and system based on graph constraints and optimal temporal transmission, aiming to solve the common problems in unstructured texts of sudden events, such as inverted causal temporal logic, difficulty in identifying implicit relationships, and unreliable reasoning due to lack of common sense constraints.

[0028] To overcome the shortcomings of existing technologies, this invention abandons traditional rule matching, simple statistical association, or purely data-driven methods that rely solely on general language models. Instead, it proposes a joint optimization framework that integrates temporal causal masking and common-sense knowledge distillation. By constructing an optimal transmission model based on temporal masking, this method introduces an infinite penalty term at the mathematical optimization level, physically blocking all reverse-temporal transmission paths, thereby ensuring that causal inference strictly conforms to the temporal preorder axiom that "the cause occurs before the effect." Simultaneously, combined with knowledge distillation technology, this invention utilizes a domain knowledge graph to construct a teacher distribution, transforming discrete physical and social common sense into a continuous prior probability distribution to guide model training. This enables accurate completion and calibration based on common-sense logic when facing implicit causality lacking explicit connectives or small-sample scenarios.

[0029] This invention can output high-quality structured causal data that is not only semantically relevant but also temporally correct and consistent with common sense logic. It effectively eliminates logical illusions and spurious correlation noise, providing a scientific and rigorous basis for emergency decision-making, risk prediction, and path blocking, and significantly improving the intelligence level and response efficiency of emergency response.

[0030] like Figure 1 As shown, the event causality mining method based on graph constraints and time-series optimal transmission proposed in this invention includes the following steps:

[0031] S11. Divide the text describing the emergency to be processed into text blocks of a preset length, and retain overlapping areas between adjacent text blocks.

[0032] In this embodiment, a raw corpus is constructed by collecting heterogeneous text data from multiple sources, including news reports, emergency plans, and social media. To address the issues of complex data sources and high noise levels, this invention employs regular expressions to remove HTML tags, garbled characters, and irrelevant advertising information. Considering that emergency event texts are typically lengthy and have broad causal logic, direct truncation could lead to the loss of crucial information. Therefore, this invention uses a sliding window mechanism to segment long emergency event descriptions into text blocks of a preset length, preserving overlapping areas between adjacent text blocks to maintain the semantic integrity of the context.

[0033] Furthermore, after segmenting the text describing the emergency into text blocks of a preset length, the method further includes: identifying absolute and relative time description information in the text blocks, converting the absolute and relative time description information into a unified timestamp format, and calculating normalized time features. Specifically, this invention introduces a key time parsing and normalization process. A time parser is used to identify absolute times (e.g., "14:00") and relative times (e.g., "later", "5 minutes later") in the text, converting them into a unified timestamp format. To adapt to subsequent neural network calculations, the system uses the Min-Max normalization method to map all timestamps to... A continuous range of values. Assume the earliest time within the current window is... The latest time is For any event occurrence time Its normalized time characteristics The calculation formula is:

[0034] ;

[0035] in, To prevent the use of tiny constants with a denominator of zero, this step ensures that physical time information is converted into numerical features that the model can directly calculate.

[0036] S12. Extract atomic events from each text block to form atomic event groups, and formalize each atomic event into a binary feature group. , Semantic content feature vectors used to characterize atomic events A timestamp used to characterize the timing of atomic events.

[0037] In this embodiment, to transform unstructured text into a machine-computable form, a sequence labeling model based on BERT-BiLSTM-CRF is used to accurately identify and extract event trigger words and their arguments (Subject, Object, Time, Location) from text blocks. Each extracted atomic event is formalized into a binary feature group. .in, It is a deep semantic vector obtained through encoding by a pre-trained language model. Specifically, the system extracts the last hidden state corresponding to the trigger word, with a dimension of [missing information]. , used to characterize the content features of an event; This is the normalized time scalar obtained in step S11, used to characterize the temporal features of events. This dual feature representation enables subsequent models to perform operations on both the semantic manifold and the temporal manifold simultaneously, which is a prerequisite for realizing temporally constrained causal mining.

[0038] S13. Combine any two atomic events in the atomic event group. and Mapped to corresponding event nodes in a pre-defined domain knowledge graph and And calculate event nodes and In the domain knowledge graph, the shortest path topological distance is used to construct a teacher model probability distribution based on the shortest path topological distance between any two atomic events. This teacher model probability distribution is used to indicate the probability of events based on human common sense. Caused the incident The prior probability.

[0039] This invention constructs a dual-path coupled causal mining model, which includes a data-driven "student network" and a knowledge-driven "teacher network".

[0040] Specifically, to address the difficulties in identifying implicit causality and the lack of common-sense constraints, the system introduces a domain knowledge graph as a teacher signal. For any two events in the input text... and The system first uses entity linking technology to map them to corresponding nodes in the domain knowledge graph. and Subsequently, the system uses breadth-first search or Dijkstra's algorithm to calculate the shortest path distance between two nodes in the graph. This distance reflects the degree of logical coherence between two events: the shorter the distance, the stronger the connection between them in terms of physical common sense. To transform discrete path distances into probability distributions usable for supervision, a teacher model probability distribution is constructed based on the topological distance of the shortest path between any two atomic events.

[0041] S14, Calculate the vector and The cosine distance between them and treat it as an event To the event The basic semantic transmission cost is determined by applying temporal constraints to it. A temporal transmission cost matrix is ​​constructed based on the temporal transmission costs of any two atomic events in an atomic event group. The Sinkhorn iterative algorithm is then used to search for the optimal transmission matrix that minimizes the total transmission cost, thus obtaining the student model's predicted probability distribution. Each element in the optimal transmission matrix represents the probability distribution of events under the premise of textual semantic similarity and temporal legitimacy. Caused the incident The confidence level.

[0042] Specifically, to address the limitation of existing optimal transmission algorithms in handling temporal logic, this invention innovatively constructs a transmission cost matrix with a "temporal causality mask". First, the cosine distance between the event vectors is calculated as the basic semantic transmission cost; the more similar the semantics, the smaller the distance. Then, a hard temporal mask term is superimposed to obtain the final temporal transmission cost. The temporal mask term is used to maximize the transmission cost when the timestamp of the result event is earlier than or equal to the timestamp of the cause event.

[0043] After obtaining the transmission cost matrix incorporating time constraints, the Sinkhorn iterative algorithm with entropy regularization is used to solve for the optimal transmission scheme. The aim is to find a non-negative transmission matrix. Minimize the total transmission cost while satisfying the edge distribution constraint. Because the cost matrix contains infinite terms, the Sinkhorn algorithm automatically compresses the probability values ​​at corresponding positions to near zero during iteration. After several iterations and convergence, the system outputs the optimal transmission scheme. Each element in the scheme This represents an event that, while simultaneously considering both textual semantic similarity and physical temporal legitimacy, It is an event The confidence score of the cause.

[0044] S15. Construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model. By minimizing the joint loss function, obtain the event causal relationship that satisfies the temporal optimal transmission under the domain knowledge graph constraint.

[0045] This invention provides an event causal mining method based on graph constraints and temporal optimal transmission. By introducing a teacher probability distribution construction mechanism based on domain knowledge graphs, it creatively transforms discrete common-sense logic into continuous probability distributions, using these as "soft labels" to guide student model training. This allows the model to no longer rely solely on word co-occurrence statistics when faced with complex text lacking explicit connectives such as "cause" or "because," but instead to complete the logic using prior physical common sense. Compared to traditional statistical association mining methods, this invention effectively corrects the logical illusions easily generated by purely data-driven models, significantly improving the recall and accuracy of implicit causal relationships. Furthermore, this invention solves the temporal logic paradox problem in causal inference from the algorithmic level. Unlike existing standard optimal transmission algorithms that only consider semantic space distance while ignoring physical time constraints, this invention innovatively introduces a temporal causal masking mechanism into the transmission cost matrix. This mechanism can forcibly block all reverse-temporal probability transmission paths, ensuring that the mined causal chains logically and naturally satisfy the "causal preorder" axiom. This guarantees the physical consistency of the output results without the need for subsequent complex rule filtering, greatly improving the credibility of the mining results.

[0046] In this embodiment of the invention, step S13, which involves constructing the teacher model probability distribution based on the shortest path topological distance between any two atomic events, specifically includes:

[0047] Based on the shortest path topological distance, a teacher model probability distribution is constructed using a Softmax function with a temperature coefficient. The calculation formula is as follows:

[0048] ;

[0049] in, Representing event nodes on a domain knowledge graph and The topological distance of the shortest path in the graph; This is a temperature coefficient used to adjust the smoothness of the distribution; The total number of events in the atomic event group. Used to indicate events based on common sense. Caused the incident The prior probability.

[0050] In this embodiment of the invention, step S14, which involves applying time constraints to the basic semantic transmission cost to obtain the time-series transmission cost, includes:

[0051] A pre-defined temporal physical constraint function, also known as a temporal mask term, is superimposed on the basic semantic transmission cost. To construct the final timing transmission cost :

[0052] ;

[0053] in, The temporal physical constraint function is defined as a piecewise function:

[0054] .

[0055] Specifically, in this formula In actual calculations, a very large constant is used instead. Through this design, the timestamp of the candidate result event... The timestamp of the causal event is earlier than or equal to the timestamp of the causal event. When the transmission cost approaches infinity, all reverse-time transmission paths are forcibly blocked at the mathematical level, ensuring that the causal relationship calculated by the model strictly follows the axiom of "causal prior order".

[0056] In this embodiment of the invention, the joint loss function for constructing the probability distribution of the teacher model and the predicted probability distribution of the student model in step S15 includes:

[0057] Constructing the optimal transmission loss based on timing masks And knowledge distillation loss used to measure the difference between the student model's predicted probabilities and the teacher model's distributed probabilities. To achieve a balance between semantic understanding and common-sense logic, the system abandons the traditional step-by-step training model and designs an end-to-end joint loss function. The function consists of two weighted parts: the first part is the optimal transmission loss based on the timing mask. The first part is used to optimize data-driven student networks, enabling them to find semantically best-matching causal pairs within the temporally valid range; the second part is knowledge distillation loss. Used to measure the transmission scheme generated by student networks. Prior distribution generated by the teacher network The differences between them.

[0058] Calculate the optimal transmission loss and knowledge distillation loss The joint loss function is as follows:

[0059]

[0060] in, This is the student model prediction probability output by the Sinkhorn algorithm; It is a teacher model distribution probability generated based on knowledge graphs; It is the numerical stability constant; These are balancing hyperparameters used to adjust the weights of common-sense constraints. During backpropagation, the gradients are passed through the Sinkhorn layer and propagated back to the BERT encoder, forcing the semantic vectors... An update has occurred. By minimizing this joint loss, the model is forced to approximate the common-sense logic of the domain as closely as possible while satisfying strict temporal constraints, thereby effectively eliminating the logical illusion caused by purely data-driven approaches.

[0061] The event causal mining method based on graph constraints and optimal temporal transmission proposed in this invention also includes a small-sample parameter adaptation step based on meta-learning. Specifically, addressing the problem of numerous types of sudden events and the scarcity of labeled data for new types, this invention introduces a model-independent meta-learning strategy as a rapid adaptation method, with the following steps:

[0062] Construct a meta-task distribution, where each meta-task contains a support set and a query set. Both the support set and the query set only need to contain a small amount of sample data.

[0063] When performing event causality mining in a new domain, the global parameters obtained after the system's pre-training are acquired. Utilizing the meta-tasks of the current domain to support the calculation of the joint loss function from a set of samples. The gradient is calculated, and one or more gradient descent updates are performed to obtain temporary parameters adapted to the current domain. :

[0064] ;

[0065] in, The learning rate for the inner loop. To support the set, For loss function The gradient.

[0066] When faced with unexpected events in new fields, this invention does not require retraining from scratch. Instead, it calculates the aforementioned joint loss function based on pre-trained global parameters θ and a small number of samples from the support set. By calculating the gradient and performing one or more gradient descent updates, temporary parameters adapted to the current domain can be obtained. Since the core logic of the model is constrained by general physical time series and common sense graphs, this transfer is essentially a semantic fine-tuning of domain-specific terms rather than relearning causal concepts, thus achieving efficient cross-domain transfer with only 5-10 samples.

[0067] The event causal mining method based on graph constraints and optimal temporal transmission proposed in this embodiment of the invention also includes the steps of causal chain topology consistency verification and decision output.

[0068] The implementation methods for causal chain topological consistency verification include:

[0069] Select students whose weights exceed a preset threshold from the student model's predicted probability distribution. The elements are used as candidate causal edges, and a global directed graph is constructed based on the candidate causal edges;

[0070] A topological sort is performed on the global directed graph to transform it into a sequence-constrained graph. Cyclic edges are detected and removed using the topological sorting algorithm to ensure that the sequence-constrained graph satisfies the acyclic property.

[0071] Although the model has incorporated timing constraints at the algorithmic level, this invention also includes a post-processing logic verification step to ensure the absolute rigor of the final output. First, the transmission scheme matrix is ​​output based on the trained student model. Filter out those with a weight higher than a preset threshold. Elements with a value of 0.6 or higher are used as candidate causal edges to construct a global directed graph. Subsequently, the system uses Karn's algorithm or depth-first search (DFS) to perform topological sorting and remove cyclic edges from the generated directed graph. The system also checks for logical loops (i.e.,...) in the graph. The system performs a secondary verification by combining the original timestamp sequence of events with the graph's structure. If any structure that violates the properties of a directed acyclic graph is found, the system will remove the edge with the lowest weight in the cycle based on the edge's confidence weight, until the graph structure meets the topological consistency requirements.

[0072] The implementation methods of decision output include:

[0073] The betweenness centrality of each node is calculated based on the sequence constraint graph. Critical blocking points on the propagation path are identified based on the betweenness centrality, and corresponding emergency blocking suggestions are generated based on these critical blocking points. The calculation formula is:

[0074] ;

[0075] in, Indicates from the source node To the target node The number of all shortest causal paths, This indicates the nodes it passes through. The number of paths.

[0076] In this embodiment, the validated causal graph is ultimately converted into standardized JSON data output, including entity information, relationship types, and confidence weights. To aid emergency decision-making, this invention further calculates the betweenness centrality of each node based on the causal graph to identify "choke points" in the propagation path, i.e., critical blocking points. The system automatically marks these points. The node with the highest value is designated as a "critical blocking point" (for example, the node "no notification received" may be connected to "base station damage" and "personnel casualties"), and corresponding emergency blocking suggestions are generated (such as "it is recommended to prioritize the use of backup communication channels"), which are then directly pushed to the command system's large screen.

[0077] The event causal mining method based on graph constraints and temporal optimal transmission proposed in this invention significantly breaks through the performance bottleneck of existing deep learning models in implicit causal mining. By introducing a teacher distribution construction mechanism based on domain knowledge graphs, this invention creatively transforms discrete common sense logic into a continuous probability distribution, serving as "soft labels" to guide model training. This design enables the model to no longer rely solely on word co-occurrence statistics when faced with complex text lacking explicit connectives such as "caused" or "because," but instead to complete the logic using prior physical common sense such as "flammable materials caused a fire." Compared to traditional statistical association mining methods, this invention effectively corrects the logical illusions easily generated by purely data-driven models, significantly improving the recall and accuracy of implicit causal relationships. Secondly, this invention solves the temporal logic paradox problem in causal inference from the algorithmic level. Unlike existing standard optimal transmission algorithms that only consider semantic space distance while ignoring physical time constraints, this invention innovatively introduces a temporal causal masking mechanism in the transmission cost matrix. By introducing an infinite penalty term during mathematical optimization, this system can forcibly block all reverse-temporal probability transmission paths. This ensures that the discovered causal chains logically and naturally satisfy the axiom of "causal priori order," guaranteeing the physical consistency of the output results without the need for subsequent complex rule filtering, thus greatly improving the credibility of the results. Finally, this invention possesses strong cross-domain transferability and small-sample adaptability. Since the core constraints of this system originate from a relatively stable physical and social common sense graph, rather than merely fitting textual features of a specific dataset, the model can quickly adapt to new emergency scenarios using general causal logic, significantly reducing its dependence on large-scale labeled data. The highly structured causal network and key node blocking suggestions ultimately output by the system can be directly applied to emergency plan formulation and risk mitigation, providing commanders with intuitive, scientific, and logically rigorous decision support, thereby significantly improving the intelligence level and response efficiency of emergency response.

[0078] The following section uses the text analysis of a highly complex emergency event, "secondary disasters caused by an earthquake in a chemical plant," as an example to describe in detail the specific implementation process of this invention in multimodal data preprocessing, common sense graph teacher distribution construction, time-series mask optimal transmission modeling, and intelligent decision support.

[0079] The following is a partial text excerpt from an emergency drill report of a chemical plant accident:

[0080] "Text: At 14:00 this afternoon, a 5.8-magnitude earthquake occurred in Park A. The tremor was strong, causing a crack in the base of storage tank No. 3. At 14:15, liquid ammonia leaked from the crack. As the wind direction in the park changed to southeast, the toxic gas rapidly spread to the nearby residential area. Due to damage to the communication base station, some residents did not receive the evacuation notice in time, resulting in several people being poisoned and falling into a coma."

[0081] I. Multimodal data preprocessing and event formalization.

[0082] Step a1: Refined Text Cleaning and Time Sequence Normalization. First, the input emergency reports in PDF or webpage format are standardized and cleaned. Using Python's regular expression library, irrelevant HTML tags, headers and footers, and special Unicode characters are removed. Considering that causal logic in emergency event texts often spans multiple sentences, a sliding window size is set to prevent semantic truncation. Each token is assigned a value, and the overlapping area between adjacent windows is set to a specific value. Each token. Simultaneously, a time parser is used to identify time information in the text. In this example, the absolute times "14:00" (earthquake occurred) and "14:15" (leakage occurred) are identified. The system uses a Min-Max normalization formula to map 14:00 on the day of the exercise to... 15:00 is mapped to Then "14:15" is converted into a normalized timestamp. This numerical processing provides the physical basis for subsequent timing mask calculations.

[0083] Step a2: Event structuring based on sequence labeling. The cleaned text stream is input into a pre-trained BERT-BiLSTM-CRF model for entity and event extraction. The model not only identifies noun entities but also focuses on extracting action trigger words and their corresponding context vectors. After inference, the unstructured text is mapped to a set of standardized atomic event tuples as follows. (Includes semantic vectors) and timestamp ):

[0084] = (Natural disaster, Park A, occurred, 5.8 magnitude earthquake,) , -);

[0085] = (Facility damage, Tank No. 3, developed, base crack) , -);

[0086] = (Hazardous chemical accident, liquid ammonia, occurred, leak,) (cracks);

[0087] = (Environmental conditions, wind direction, shifted to, southeast wind,) , -);

[0088] = (Secondary disasters, toxic gases, diffusion, -, (residential area);

[0089] = (Facility damage, Communication base station, -, Damaged) , -);

[0090] = (Management deficiencies, residents, did not receive, evacuation notice) , -);

[0091] = (Casualties, Multiple people, -, Poisoning and coma, , -);

[0092] II. Construction of Teacher Distribution Based on Common Sense Graph

[0093] Step b1: Entity Linking and Graph Path Retrieval. To address the difficulty of implicit relationship identification, this invention introduces a "knowledge graph of the emergency event domain" as a teacher signal. First, through entity linking technology, the extracted events are mapped to graph nodes. For example, [the following is a separate section, likely related to event identification]. (Base station damaged) is mapped to graph node ID: 201, and will (Poisoning and coma) is mapped to node ID: 509. This applies to text that is not directly linked. and Purely statistical models struggle to establish connections. However, within a knowledge graph, the system retrieved a clear logical path: [base station]. [Communication Network] [Warning Information] [Personnel evacuation] [Life safety]

[0094] Step b2: Soft label generation and teacher distribution construction. Calculate the topological distance of the above paths in the graph. Because the logical link is clear, the distance is relatively short. The system utilizes a Softmax function with a temperature coefficient; optionally, the temperature coefficient... Transform this distance into a teacher probability distribution In this distribution, The value was calculated to be 0.85, meaning that from a common-sense perspective, "damage to communication facilities" is highly likely to lead to "personnel safety accidents." This high-confidence "soft label" will serve as a supervisory signal, forcing student models to pay attention to this implicit association during subsequent training, thus avoiding random guessing like "blind men and the elephant."

[0095] III. Construction of the optimal transmission model based on timing mask.

[0096] Step c1: Cost calculation for semantic and temporal fusion. The system constructs a data-driven student model and calculates... Transmission cost matrix .by (Poisoning, t=0.40) and Taking an earthquake (t=0.0) as an example, the two are highly semantically correlated, with a basic semantic distance of only 0.3. In this case, without constraints, traditional models are highly prone to misjudging it as "poisoning caused the earthquake." This invention introduces a core temporal causal mask here. The system checked the timestamps and found... If it conforms to logic, the mask is 0; otherwise, in calculating from point to During reverse transmission, due to violation of causal preorder, the mask value is forcibly set to 0. Ultimately, the matrix The reverse timing position was filled with infinity, physically blocking the erroneous path.

[0097] Step c2: Sinkhorn Algorithm Solution and Alignment. The cost matrix incorporating the timing mask is input into the Sinkhorn iterative algorithm. After approximately 20 iterations and convergence, the system calculates the optimal transmission scheme. The results show that although "poisoning" and "earthquake" are semantically similar, the transmission weight is suppressed to 0 by the timing mask. Conversely, "earthquake" The weight of "leakage" is 0.92, and "base station damage" is also considered. No notification received The weight of "poisoning" was accurately identified and increased to 0.88 under the calibration of teacher distribution.

[0098] IV. Joint Optimization and Decision Support.

[0099] Step d1: Model validation under dual constraints. The system employs a joint loss function during inference. Verification is required. On one hand, the optimal transmission loss... This ensures that the generated causal chain flows sequentially along the timeline (from 14:00 to 14:15); on the other hand, KL divergence loss ensures that the mined implicit relationships conform to common-sense logic in the graph. After passing both verifications, the system confirms that the causal graph possesses both physical and logical self-consistency.

[0100] Step d2: Structured Causal Network Output. The system ultimately generates the following standardized JSON causal mining report, including event entities, relationship types, weights, and remarks. Note: Thanks to time-series masking, all reverse causal relationships are automatically removed from the results:

[0101] { "EventEntities": ["Earthquake", "Base Crack", "Liquid Ammonia Leak", "Southeast Wind", "Gas Dispersion", "Base Station Damage", "No Notification Received", "Personnel Poisoning"], "CausalGraph": [ {"Source": "Earthquake", "Target": "Base Crack", "Type": "Explicit", "Weight": 0.98}, {"Source":"Base Crack", "Target": "Liquid Ammonia Leak", "Type": "Explicit", "Weight": 0.96}, {"Source": "Liquid Ammonia Leak", "Target": "Gas Dispersion", "Type": "Explicit", "Weight":0.95}, {"Source": "Southeast Wind", "Target": "Gas Dispersion", "Type": "Implicit", "Weight": 0.88, "Note": "Physical Constraints"}, {"Source": "Earthquake", "Target": "Base Station Damage", "Type": "Explicit", "Weight": 0.92}, {"Source": "Base Station Damage", "Target": "No Notification Received", "Type": "Implicit-CommonSense", "Weight": 0.94}, {"Source": "No notification received", "Target": "Personnel poisoned", "Type": "Indirect", "Weight": 0.90}, ]}.

[0102] Step d3: Centrality-based decision-making blocking recommendations. Based on the generated causal network, the system automatically calculates the betweenness centrality of each node. Analysis reveals two high-centrality nodes in the causal graph. Based on this, the system generates specific blocking recommendations: 1. Source blocking strategy: targeting high-centrality nodes. (Leakage) It is recommended to immediately dispatch chemical defense troops to seal the cracks and cut off the source of the toxic substance. 2. Path blocking strategy: Target key intermediary nodes. (Without notification), the system detected that this was a high-weight path leading to casualties and originated from base station damage. Therefore, the decision-making system recommends immediately activating "non-communication-dependent" physical alarm methods (such as vehicle-mounted air raid sirens or drone-borne announcements) to cover the blind spot, forcibly cutting off the evolution path from "information deficiency" to "personnel casualties," and minimizing secondary disaster losses.

[0103] For the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0104] This invention provides an event causal mining system based on graph constraints and time-series optimal transmission. The system includes functional modules for implementing the event causal mining method based on graph constraints and time-series optimal transmission as described in any of the preceding claims. Figure 2 The schematic diagram illustrates the structure of the event causality mining system based on graph constraints and time-series optimal transmission provided in an embodiment of the present invention. (Refer to...) Figure 2 The system described in this embodiment of the invention includes:

[0105] The preprocessing module 201 is used to divide the text describing the sudden event to be processed into text blocks of a preset length, and to retain overlapping areas between adjacent text blocks;

[0106] Event extraction module 202 is used to extract atomic events from each text block to form atomic event groups, and to formalize each atomic event into a binary feature group. , Semantic content feature vectors used to characterize atomic events Timestamps used to characterize the timing of atomic events;

[0107] Common Sense Graph Causal Mining Module 203 is used to extract any two atomic events from an atomic event group. and Mapped to corresponding event nodes in a pre-defined domain knowledge graph and And calculate event nodes and In the domain knowledge graph, the shortest path topological distance is used to construct a teacher model probability distribution based on the shortest path topological distance between any two atomic events. This teacher model probability distribution is used to indicate the probability of events based on human common sense. Caused the incident The prior probability;

[0108] The temporal causal mining module 204 is used to calculate vectors. and The cosine distance between them and treat it as an event To the event The basic semantic transmission cost is obtained by applying time constraints to the basic semantic transmission cost. The time transmission cost matrix is ​​constructed based on the time transmission costs of any two atomic events in the atomic event group. The optimal transmission matrix that minimizes the total transmission cost is searched using the Sinkhorn iterative algorithm to obtain the student model prediction probability distribution.

[0109] The joint optimization module 205 is used to construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model. By minimizing the joint loss function, the event causal relationship that satisfies the temporal optimal transmission under the domain knowledge graph constraint is obtained.

[0110] In this embodiment of the invention, the system further includes a domain adaptation module (not shown in the accompanying drawings). This domain adaptation module is used to construct a meta-task distribution, where each meta-task includes a support set and a query set. When performing event causal mining on a new domain, it obtains the global parameters pre-trained by the system. Utilizing the meta-tasks of the current domain to support the calculation of the joint loss function from a set of samples. The gradient is calculated, and one or more gradient descent updates are performed to obtain temporary parameters adapted to the current domain. :

[0111] ;

[0112] in, The learning rate for the inner loop. To support the set, For loss function The gradient.

[0113] In this embodiment of the invention, the system further includes a logic verification module and a decision output module (not shown in the accompanying drawings), wherein:

[0114] The logic verification module is used to filter out students whose weights exceed a preset threshold from the student model's predicted probability distribution. The elements are used as candidate causal edges, and a global directed graph is constructed based on the candidate causal edges. The global directed graph is topologically sorted to transform it into a sequence constraint graph. Cyclic edges are detected and removed by the topological sorting algorithm to ensure that the sequence constraint graph satisfies the acyclicity property.

[0115] The decision output module calculates the betweenness centrality of each node based on the sequence constraint graph, identifies key blocking points on the propagation path based on the betweenness centrality, and generates corresponding emergency blocking suggestions based on these key blocking points. The calculation formula is:

[0116] ;

[0117] in, Indicates from the source node To the target node The number of all shortest causal paths, This indicates the nodes it passes through. The number of paths.

[0118] As the system implementation is basically similar to the method implementation, the description is relatively simple, and relevant parts can be found in the description of the method implementation.

[0119] In addition, another embodiment of the present invention provides a computer program product on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps in the above embodiment of the event causal mining method based on graph constraints and time-series optimal transmission.

[0120] Furthermore, another embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps described above in the embodiment of the event causal mining method based on graph constraints and time-series optimal transmission.

[0121] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, any of the claimed embodiments can be used in any combination.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for mining event causality based on graph constraints and optimal temporal transmission, characterized in that, The method includes: The text describing the emergency to be processed is divided into text blocks of a preset length, and overlapping areas are preserved between adjacent text blocks; Atomic events are extracted from each text block to form atomic event groups, and each atomic event is formalized into a binary feature group. , Semantic content feature vectors used to characterize atomic events Timestamps used to characterize the timing of atomic events; Any two atomic events in the atomic event group and Mapped to corresponding event nodes in a pre-defined domain knowledge graph. and And calculate event nodes and In the domain knowledge graph, the shortest path topological distance is used to construct a teacher model probability distribution based on the shortest path topological distance between any two atomic events. This teacher model probability distribution is used to indicate the probability of events based on human common sense. Caused the incident The prior probability; Calculate vectors and The cosine distance between them and treat it as an event To the event The basic semantic transmission cost is obtained by applying time constraints to the basic semantic transmission cost. The time transmission cost matrix is ​​constructed based on the time transmission costs of any two atomic events in the atomic event group. The optimal transmission matrix that minimizes the total transmission cost is searched using the Sinkhorn iterative algorithm to obtain the student model prediction probability distribution. We construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model. By minimizing the joint loss function, we obtain the event causal relationship that satisfies the temporal optimal transmission under the domain knowledge graph constraint.

2. The method according to claim 1, characterized in that, Construct the teacher model probability distribution based on the topological distance of the shortest path corresponding to any two atomic events, including: Based on the shortest path topological distance, a teacher model probability distribution is constructed using a Softmax function with a temperature coefficient. The calculation formula is as follows: ； in, Representing event nodes on a domain knowledge graph and The topological distance of the shortest path in the graph; This is a temperature coefficient used to adjust the smoothness of the distribution; The total number of events in the atomic event group. Used to indicate events based on common sense. Caused the incident The prior probability.

3. The method according to claim 1, characterized in that, The temporal transmission cost is obtained by applying temporal constraints to the basic semantic transmission cost, including: A pre-defined temporal physical constraint function is superimposed on the basic semantic transmission cost to construct the final temporal transmission cost. : ； in, The temporal physical constraint function is defined as a piecewise function: 。 4. The method according to claim 1, characterized in that, Construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model, including: Constructing the optimal transmission loss based on timing masks And knowledge distillation loss used to measure the difference between the student model's predicted probabilities and the teacher model's distributed probabilities. ; Calculate the optimal transmission loss and knowledge distillation loss The joint loss function is as follows: in, This is the student model prediction probability output by the Sinkhorn algorithm; It is a teacher model distribution probability generated based on knowledge graphs; It is the numerical stability constant; It is a balancing hyperparameter; This incurs the cost of timing transmission.

5. The method according to claim 1, characterized in that, The method further includes: Construct a meta-task distribution, where each meta-task contains a support set and a query set; When performing event causal mining in a new domain, the global parameters obtained after the system's pre-training are acquired. Utilizing the meta-tasks of the current domain to support the calculation of the joint loss function from a set of samples. The gradient is calculated, and one or more gradient descent steps are performed to update the parameters, resulting in temporary parameters adapted to the current domain. : ； in, The learning rate for the inner loop. To support the set, loss function The gradient.

6. The method according to claim 1, characterized in that, After segmenting the text describing the emergency to be processed into text blocks of a preset length, the method further includes: Identify absolute and relative time descriptions in text blocks, convert them into a unified timestamp format, and calculate normalized time features.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Select students whose weights exceed a preset threshold from the student model's predicted probability distribution. The elements are used as candidate causal edges, and a global directed graph is constructed based on the candidate causal edges; A topological sort is performed on the global directed graph to transform it into a sequence-constrained graph. Cyclic edges are detected and removed using the topological sorting algorithm to ensure that the sequence-constrained graph satisfies the acyclic property. The betweenness centrality of each node is calculated based on the sequence constraint graph. Critical blocking points on the propagation path are identified based on the betweenness centrality, and corresponding emergency blocking suggestions are generated based on these critical blocking points. The calculation formula is: ； in, Indicates from the source node To the target node The number of all shortest causal paths, This indicates that it passes through the nodes. The number of paths.

8. An event causal mining system based on graph constraints and time-series optimal transmission, characterized in that, The system includes: The preprocessing module is used to divide the text describing the emergency into text blocks of a preset length, while preserving overlapping areas between adjacent text blocks; The event extraction module is used to extract atomic events from each text block to form atomic event groups, and to formalize each atomic event into a binary feature group. , Semantic content feature vectors used to characterize atomic events Timestamps used to characterize the timing of atomic events; The common sense graph causal mining module is used to extract causal information from any two atomic events in an atomic event group. and Mapped to corresponding event nodes in a pre-defined domain knowledge graph. and And calculate event nodes and In the domain knowledge graph, the shortest path topological distance is used to construct a teacher model probability distribution based on the shortest path topological distance between any two atomic events. This teacher model probability distribution is used to indicate the probability of events based on human common sense. Caused the incident The prior probability; The temporal causal mining module is used to calculate vectors. and The cosine distance between them and treat it as an event To the event The basic semantic transmission cost is obtained by applying time constraints to the basic semantic transmission cost. The time transmission cost matrix is ​​constructed based on the time transmission costs of any two atomic events in the atomic event group. The optimal transmission matrix that minimizes the total transmission cost is searched using the Sinkhorn iterative algorithm to obtain the student model prediction probability distribution. The joint optimization module is used to construct a joint loss function for the probability distribution of the teacher model and the predicted probability distribution of the student model. By minimizing the joint loss function, the event causal relationship that satisfies the temporal optimal transmission under the constraints of the domain knowledge graph is obtained.

9. The system according to claim 8, characterized in that, The system also includes: The domain adaptation module is used to construct a meta-task distribution. Each meta-task contains a support set and a query set. When performing event causal mining on a new domain, it obtains the global parameters pre-trained by the system. Utilizing the meta-tasks of the current domain to support the calculation of the joint loss function from a set of samples. The gradient is calculated, and one or more gradient descent steps are performed to update the parameters, resulting in temporary parameters adapted to the current domain. : ； in, The learning rate for the inner loop. To support the set, loss function The gradient.

10. The system according to claim 8, characterized in that, The system also includes: The logic verification module is used to filter out students whose weights exceed a preset threshold from the student model's predicted probability distribution. The elements are used as candidate causal edges, and a global directed graph is constructed based on the candidate causal edges. The global directed graph is topologically sorted to transform it into a sequence constraint graph. Cyclic edges are detected and removed by the topological sorting algorithm to ensure that the sequence constraint graph satisfies the acyclicity property. The decision output module calculates the betweenness centrality of each node based on the sequence constraint graph, identifies key blocking points on the propagation path based on the betweenness centrality, and generates corresponding emergency blocking suggestions based on these key blocking points. The calculation formula is: ； in, Indicates from the source node To the target node The number of all shortest causal paths, This indicates that it passes through the nodes. The number of paths.

Images