A Method and System for Constructing an Event Chain Ontology Map of Urban Rainstorm Flood Disasters

Abstract

The present application belongs to the technical field of knowledge graph construction, and discloses a method and system for constructing a cause-and-effect graph of urban rainstorm and flood disaster chain, which includes: first obtaining research literature, first historical events and secondary events, constructing a directed network graph of the disaster chain and constructing an ontology model based on it; then obtaining the second historical event and secondary events, pre-processing and inputting them into the ontology model to obtain an event corpus; extracting event trigger words from the corpus using the multimodal dynamic fusion model MDF‑Net to build a database; extracting events in the corpus based on the trigger words to obtain event types and arguments; then using the improved ELECTRA‑BiLSTM model to extract the time sequence, spatial relationship and semantic relationship of the trigger words; finally, constructing a cause-and-effect graph of the disaster chain with event types, arguments, time sequence and spatial relationship as nodes and semantic relationship as edges. This method can effectively integrate multi-source data and complex disaster information, and improve the integrity and accuracy of the cause-and-effect graph of urban rainstorm and flood disaster chain.

Description

Technical Field

[0001] The present application belongs to the technical field of knowledge graph construction, and specifically relates to a method and system for constructing a knowledge graph of urban rainstorm and flood disaster chain. Background Art

[0002] As an important branch of knowledge graph, the dynamic deduction and causal reasoning characteristics of the event graph provide a new technical path for disaster chain modeling. Among them, the rainstorm and flood disaster chain has typical spatiotemporal coupling characteristics.

[0003] The current research on disaster chain maps mainly adopts three types of technical routes: the first is disaster event association mining based on statistical learning, which uses association rule mining algorithms to extract high-frequency co-occurring disaster elements; the second is an event extraction model based on deep learning, which realizes disaster entity recognition through pre-training models such as BERT; the third is a disaster deduction method based on complex network theory, which constructs a topological network in which nodes represent disaster events and edges represent evolutionary relationships.

[0004] However, the existing methods for constructing urban rainstorm and flood disaster chain maps have many shortcomings. On the one hand, most methods fail to fully consider the spatiotemporal characteristics of urban rainstorm and flood disaster chains. In the actual process of disaster evolution, temporal and spatial factors play a key role in the development and propagation of disaster events, but the existing maps are difficult to accurately reflect these characteristics, resulting in the inability of the constructed maps to truly and comprehensively present the actual evolution process of the disaster chain. On the other hand, the existing map construction methods have obvious limitations in data processing, event extraction, and relationship identification. These methods pay more attention to the primary events of disasters, but pay insufficient attention to the secondary disaster events caused by the primary disaster events, resulting in a large amount of important disaster information being ignored. This leads to the inability to effectively integrate multi-source data and complex disaster information, which in turn affects the integrity and accuracy of the map, and it is difficult to meet the needs of actual disaster research and management. Therefore, there is an urgent need for a method for constructing an urban rainstorm and flood disaster chain map that can fully take into account the spatiotemporal characteristics and effectively integrate multi-source data. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a method and system for constructing a map of urban rainstorm and flood disaster chain events, which can solve the problem that the existing methods for constructing urban rainstorm and flood disaster chain maps do not fully take into account the temporal and spatial characteristics and effectively integrate multi-source data.

[0006] In order to solve the above technical problems, this application is implemented as follows:

[0007] In a first aspect, an embodiment of the present application provides a method for constructing a chain diagram of urban rainstorm and flood disasters, the method comprising:

[0008] Obtain research literature and first historical events and their secondary events of rainstorm and flood disasters in the target city, and construct a directed network diagram of the disaster chain;

[0009] According to the directed network diagram of the disaster chain, an ontology model of the rainstorm and flood disaster in the target city is constructed;

[0010] The second historical event and its secondary events of the rainstorm and flood disaster in the target city are obtained for preprocessing, and the preprocessed second historical event and its secondary events are input into the ontology model to obtain an event corpus;

[0011] Using a multimodal dynamic fusion model MDF-Net to extract event trigger words from the event corpus to construct an event trigger word library;

[0012] According to the trigger words in the event trigger word library, event extraction is performed in the event corpus to obtain the event type and event argument corresponding to the trigger words;

[0013] According to the trigger words in the event trigger vocabulary, performing relationship extraction in the event corpus to obtain semantic relationships, temporal relationships and spatial relationships related to the trigger words, wherein the temporal relationships and the spatial relationships are obtained by an improved ELECTRA-BiLSTM model, and the improved ELECTRA-BiLSTM model includes an ELECTRA model, a hole convolution layer, a gating mechanism layer, a BiLSTM model and a multi-head attention mechanism layer;

[0014] A disaster chain event graph of rainstorm and flood disasters in the target city is constructed by taking the event type, the event argument, the temporal relationship and the spatial relationship as event nodes and taking the semantic relationship as the edge of the event.

[0015] As an optional implementation manner of the first aspect of the present application, the steps of obtaining research literature and a first historical event and its secondary events on rainstorm and flood disasters in a target city, and constructing a directed network diagram of a disaster chain include: obtaining research literature on rainstorm and flood disasters in the target city, and organizing the research literature into a research review using a literature research method; obtaining the first historical event and its secondary events on rainstorm and flood disasters in the target city, and organizing the first historical event and its secondary events into case events using a case analysis method; constructing a disaster chain event set based on the research review and the case events; and drawing a directed network diagram of a disaster chain corresponding to the rainstorm and flood disasters in the target city based on the disaster chain event set.

[0016] As an alternative implementation of the first aspect of the present application, the steps of using the multi-modal dynamic fusion model MDF-Net to extract event trigger words from the event corpus include: performing multi-granularity annotation on the unstructured text of the event corpus, adopting the BIOES annotation system and combining with disaster domain ontology knowledge, and generating fine-grained annotation data through entity boundary enhancement and semantic role annotation; performing spatio-temporal alignment on the text data of the event corpus with the real-time meteorological data, geographic information system layer and social media public opinion data of the target city to generate text alignment data; using the cross-modal attention mechanism of the multi-modal dynamic fusion model MDF-Net to extract the correlation weights of text keywords and spatio-temporal features of the fine-grained annotation data and the text alignment data to generate a unified multi-modal embedding representation; inputting the multi-modal embedding representation into the Transformer-GAT-CRF hybrid network model, the pre-training layer of the Transformer-GAT-CRF hybrid network model adopts the ELECTRA model enhanced by the disaster domain, and generates context-aware word vectors through masked language modeling and causal prediction tasks; the graph attention layer GAT of the Transformer-GAT-CRF hybrid network model embeds the disaster ontology knowledge graph, captures the topological dependence relationship between event trigger words, and assigns weights to event, location, and time nodes through multi-head attention; the temporal-spatial bidirectional gating layer of the Transformer-GAT-CRF hybrid network model combines BiLSTM and dilated temporal convolutional network DTCN to model the temporal evolution and spatial diffusion characteristics of events respectively; the contrastive learning module of the Transformer-GAT-CRF hybrid network model optimizes the discrimination of trigger word representations by constructing positive and negative sample pairs; the CRF decoder of the Transformer-GAT-CRF hybrid network model outputs event trigger words.

[0017] As an alternative implementation of the first aspect of the present application, the steps of performing event extraction in the event corpus according to the trigger words in the event trigger word library to obtain the event types and event arguments corresponding to the trigger words include: converting the trigger words in the event trigger word library into trigger word feature vectors, clustering the trigger word feature vectors to obtain preliminary event types, and aligning the preliminary event types to obtain the final event types; converting the argument words in the event trigger word library into argument word feature vectors, clustering the argument word feature vectors to obtain preliminary event arguments, and aligning the preliminary event arguments to obtain the final event arguments.

[0018] As an optional implementation manner of the first aspect of the present application, the step of aligning the preliminary event types to obtain the final event types includes: calculating a first cosine similarity between any two event types, setting a first threshold, and determining whether the first cosine similarity is greater than or equal to the first threshold; if so, the two event types are regarded as the same event type; the step of aligning the preliminary event arguments to obtain the final event arguments includes: calculating a second cosine similarity between any two event arguments, setting a second threshold, and determining whether the second cosine similarity is greater than or equal to the second threshold; if so, the two event arguments are regarded as the same event argument.

[0019] As an optional implementation of the first aspect of the present application, according to the trigger words in the event trigger vocabulary, relationship extraction is performed in the event corpus, and the step of obtaining semantic relationships related to the trigger words includes: selecting text fragments of multiple trigger words in the event trigger vocabulary to form candidate event pairs; after word segmentation and part-of-speech tagging of the candidate event pairs, constructing a dependency syntax tree to analyze the preliminary semantic relationships between words; based on the preliminary semantic relationships, establishing a mapping rule library from specific to abstract events; performing logical deduction and verification on the mapping rule library to obtain a dependency pattern rule library; and adjusting the trigger word weights in the dependency pattern rule library for optimization to obtain extraction results containing semantic relationships.

[0020] As an optional implementation of the first aspect of the present application, the process of obtaining the temporal relationship and the spatial relationship by the improved ELECTRA-BiLSTM model includes: selecting text fragments of multiple trigger words in the event trigger vocabulary, inputting the pre-trained text encoder ELECTRA model for pre-training, and obtaining the corresponding text word vector sequence; inputting the text word vector sequence into the void convolution layer to extract the local-global features of the text; inputting the local-global features into the BiLSTM model to capture the context dependency; inputting the context dependency into the multi-head attention mechanism layer, and calculating the first attention of the temporal task and the second attention of the spatial task respectively; and obtaining the classification prediction results of the temporal relationship and the spatial relationship according to the first attention and the second attention respectively.

[0021] In a second aspect, the embodiment of the present application provides a system for constructing a chain diagram of urban rainstorm and flood disasters, the system comprising:

[0022] The data acquisition module is used to obtain research literature and the first historical events and their secondary events of rainstorm and flood disasters in the target city, and to construct a directed network diagram of the disaster chain;

[0023] An ontology modeling module, used to construct an ontology model of the rainstorm and flood disaster in the target city according to the directed network diagram of the disaster chain;

[0024] A corpus generation module is used to obtain the second historical event and its secondary events of the rainstorm and flood disaster in the target city for preprocessing, and input the preprocessed second historical event and its secondary events into the ontology model to obtain an event corpus;

[0025] A trigger word extraction module, used for extracting event trigger words from the event corpus using a multimodal dynamic fusion model MDF-Net to construct an event trigger word library;

[0026] An event extraction module, used to extract events from the event corpus according to the trigger words in the event trigger word library, and obtain the event type and event argument corresponding to the trigger words;

[0027] A relationship extraction module, used for performing relationship extraction in the event corpus according to the trigger words in the event trigger vocabulary, and obtaining semantic relations, temporal relations and spatial relations related to the trigger words, wherein the temporal relations and the spatial relations are obtained by an improved ELECTRA-BiLSTM model, and the improved ELECTRA-BiLSTM model includes an ELECTRA model, a hole convolution layer, a gating mechanism layer, a BiLSTM model and a multi-head attention mechanism layer;

[0028] The graph construction module is used to construct a disaster chain event graph of the rainstorm and flood disaster in the target city with the event type, the event argument, the temporal relationship and the spatial relationship as event nodes and the semantic relationship as the edge of the event.

[0029] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect.

[0030] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0031] Compared with the prior art, this application has the following beneficial effects:

[0032] First, we obtain research literature, first historical events, and secondary events to construct a directed network diagram of the disaster chain, which can sort out the relationship between disasters and their secondary events from a macro perspective and provide a framework for subsequent modeling. Then, we construct an ontology model based on the directed network diagram to formally describe disaster-related concepts, attributes, and relationships, unify the semantic framework, and facilitate the processing of multi-source data. After obtaining the second historical event preprocessing, we input it into the ontology model to obtain the event corpus, remove noise and improve data quality, and provide high-quality data for subsequent extraction. The multimodal dynamic fusion model MDF-Net is used to extract event trigger words from the corpus to build a vocabulary, and accurately locate event clues based on the advantages of the model. Event extraction is performed based on the trigger words to obtain event types and arguments, clarify the specific content and elements of the event, and provide detailed event information for the map. The improved ELECTRA-BiLSTM model is used to extract the semantics, temporal and spatial relationships of trigger words, and a variety of neural network layers are integrated to comprehensively mine event associations. Finally, a disaster chain graph is constructed with event types, arguments, and spatiotemporal relationships as nodes and semantic relationships as edges to comprehensively and accurately describe the complex relationship between disasters and their secondary events, fully taking into account spatiotemporal characteristics and integrating multi-source data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of a method for constructing a chain diagram of urban rainstorm and flood disasters provided by the first embodiment of the present application;

[0034] Figure 2 It is a structural schematic diagram of a system for constructing a chain event graph of urban rainstorm and flood disasters provided in the second embodiment of the present application. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0036] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here. In addition, the "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally represents that the objects associated with each other are in an "or" relationship.

[0037] In order to illustrate the technical solution described in this application, a specific embodiment is provided below for illustration.

[0038] Example 1

[0039] See also Figure 1 , which is a flow chart of a method for constructing a chain event graph of urban rainstorm and flood disasters proposed in the first embodiment of the present application. The proposed method includes S01 to S07.

[0040] Step S01: Obtain research literature and the first historical event and its secondary events of rainstorm and flood disasters in the target city, and construct a directed network diagram of the disaster chain.

[0041] In some embodiments, research literature on rainstorm and flood disasters in the target city is obtained, and the research literature is organized into a research review using a literature research method; the first historical event and its secondary events of the rainstorm and flood disasters in the target city are obtained, and the first historical event and its secondary events are organized into case events using a case analysis method; a disaster chain event set is constructed based on the research review and the case events; and a disaster chain directed network diagram corresponding to the rainstorm and flood disasters in the target city is drawn based on the disaster chain event set.

[0042] Specifically, first, research literature collection and review: collect research literature on disaster and disaster chain theory, urban rainstorm and flood disaster formation mechanism, disaster risk assessment, disaster risk analysis, and natural disaster emergency management plan. Use literature research method to sort out these documents and form a review of urban rainstorm and flood disaster chain research. At the same time, list the relevant norms and laws and regulations on urban flood disasters.

[0043] Furthermore, case data collection and analysis: collect case data of rainstorm and flood disaster chain in target cities. The selected cases should be representative, typical and relevant, and the data should be authentic, reliable, detailed and sufficient. Use case analysis method to sort out the first historical events and secondary events of rainstorm and flood disasters in target cities, analyze the evolution process of disasters in detail, and organize them into case events.

[0044] Furthermore, a disaster chain event set is constructed: the disaster events that occurred are listed by combining the research review formed by literature research and the case events obtained by case analysis. According to the principle of comprehensive and concise summary of disaster events, the final disaster chain disaster event set is constructed.

[0045] Furthermore, draw a directed network diagram of the disaster chain: based on the constructed disaster chain event set, draw a directed network diagram of the disaster chain corresponding to the rainstorm and flood disasters in the target city to provide a reference for the subsequent construction of the disaster chain event map.

[0046] In this step, by obtaining research literature, historical events, and their secondary events, data from different sources and of different types are integrated. This integration makes the originally diverse and scattered data form an organic whole, providing a comprehensive and systematic data foundation for subsequent analysis. Moreover, this basic data is applicable to the target city situation of the research, has a certain regional nature, and makes the research data more accurate. The directed network diagram of the disaster chain can reflect the dynamic associations between disaster events. The rainstorm and flood disaster does not occur in isolation but will trigger a series of secondary events over time. By constructing a directed network diagram, the development process from the primary disaster to the secondary disaster is recorded, reflecting the dynamic changes in the data. This helps researchers understand the evolution law of disasters and predict possible future disaster events and their development trends.

[0047] Step S02: Construct an ontology model for the rainstorm and flood disaster in the target city based on the directed network diagram of the disaster chain.

[0048] Specifically, based on Step S01, the hierarchical relationship of the urban rainstorm and flood disaster chain is further clarified, covering the related concepts of the urban rainstorm and flood disaster chain, the attributes of disaster events, and the relationships between disaster events. Among them, the related concepts include related concepts of disaster-causing events, disaster-bearing events, and disaster-relief events. The attributes of disaster events include the time, location, participants, and objects of disaster events. The relationships between disaster events mainly include semantic relationships, temporal relationships, and spatial relationships between disaster events.

[0049] Furthermore, according to disaster theory, based on the description of the disaster chain , where is the urban rainstorm and flood disaster chain, is the disaster-forming environment, is the disaster-causing factor, is the disaster-bearing body, determine the knowledge domain boundary of the urban rainstorm and flood disaster chain, form the pattern layer of the urban rainstorm and flood disaster chain, and achieve a unified description of the urban flood disaster chain.

[0050] Furthermore, use the ontology modeling software Protégé to construct the ontology model of the urban rainstorm and flood disaster chain in a semi-automated manner.

[0051] In this step, by using the ontology modeling software Protégé, an ontology model of the urban rainstorm and flood disaster chain is constructed. This model not only clearly defines the related concepts and attributes in the disaster chain, such as disaster-causing events, disaster-bearing events, and disaster-relief events, but also includes attributes such as the time, location, participants, and objects of disaster events. In addition, it clarifies the semantic, temporal, and spatial relationships between disaster events, provides a clear definition for the knowledge domain boundary of the disaster chain, and forms a pattern layer for the unified description of the urban flood disaster chain, thus providing an application basis for disaster management and research.

[0052] Step S03: Obtain the second historical events of the target city's rainstorm and flood disasters and their secondary events for preprocessing, and input the preprocessed second historical events and their secondary events into the ontology model to obtain an event corpus.

[0053] Specifically, in the data acquisition stage, for example, write a collection program in a programming language to collect news reports on urban rainstorm and flood disaster events and their secondary events from the websites of emergency management departments and large news media websites.

[0054] In the data preprocessing stage, it includes data cleaning and word segmentation and annotation. For example, use regular expressions and manual review and other methods to preprocess redundant values and missing values in the collected data. Use a Chinese word segmentation library to perform word segmentation, part-of-speech tagging, and syntactic analysis on the collected content, delete stop words and irrelevant characters, and use regular expressions to screen news reports and delete irrelevant content.

[0055] In this step, the second historical events of the target city's rainstorm and flood disasters and their secondary events are obtained from the emergency management department and large news media websites by writing a collection program, and the collected data is preprocessed, including data cleaning, word segmentation and annotation, and syntactic analysis, effectively removing redundant values and irrelevant content, and ensuring the accuracy and consistency of the data. Finally, the preprocessed event data is input into the ontology model to construct a high-quality event corpus, providing a reliable data basis for subsequent analysis and research.

[0056] Step S04: Use the multi-modal dynamic fusion model MDF-Net to extract event trigger words from the event corpus to construct an event trigger word library.

[0057] In some embodiments, perform multi-granularity annotation on the unstructured text of the event corpus, adopt the BIOES annotation system and combine the ontology knowledge in the disaster field, and generate fine-grained annotation data through entity boundary enhancement and semantic role annotation;

[0058] Align the text data of the event corpus with the real-time meteorological data, geographical information system layers, and social media public opinion data of the target city in space and time to generate text alignment data;

[0059] Use the cross-modal attention mechanism of the multi-modal dynamic fusion model MDF-Net to extract the correlation weights of the text keywords and spatio-temporal features of the fine-grained annotation data and the text alignment data, and generate a unified multi-modal embedding representation;

[0060] The multi-modal embedding representation is input into the Transformer-GAT-CRF hybrid network model. The pre-training layer of the Transformer-GAT-CRF hybrid network model uses the ELECTRA model enhanced in the disaster domain to generate context-aware word vectors through masked language modeling and causal prediction tasks. The graph attention layer GAT of the Transformer-GAT-CRF hybrid network model embeds the disaster ontology knowledge graph, captures the topological dependencies between event trigger words, and assigns weights to event, location, and time nodes through multi-head attention. The temporal-spatial bidirectional gating layer of the Transformer-GAT-CRF hybrid network model combines BiLSTM and dilated temporal convolutional network (DTCN) to model the temporal evolution and spatial diffusion characteristics of events respectively. The contrastive learning module of the Transformer-GAT-CRF hybrid network model optimizes the discriminability of trigger word representations by constructing positive and negative sample pairs. The CRF decoder of the Transformer-GAT-CRF hybrid network model outputs event trigger words.

[0061] In this step, first, through multi-modal data integration, this step breaks through the limitation of a single modality. The model can fuse text data, real-time meteorological data of the target city, geographical information system (GIS) layers, and social media sentiment data to generate a unified multi-modal embedding representation. This integration of multi-modal data enables the model to capture various features and context information of events more comprehensively. Second, the model fully considers the spatio-temporal characteristics of events. Through spatio-temporal alignment, the model can capture the spatio-temporal correlations between text data and real-time meteorological, GIS layer, and social media sentiment data to generate text-aligned data. This helps the model better understand the spatio-temporal evolution and diffusion patterns of events. Third, the application of the Transformer-GAT-CRF hybrid network model further improves the accuracy of event trigger word extraction. This model combines pre-training, graph attention mechanism, temporal-spatial bidirectional gating layer, contrastive learning module, and CRF decoder, and can more effectively capture the topological dependencies between event trigger words and model the temporal evolution and spatial diffusion characteristics of events. In addition, the pre-training layer of the model uses the ELECTRA model enhanced in the disaster domain to generate context-aware word vectors through masked language modeling and causal prediction tasks, further improving the semantic understanding ability of the model. Finally, the implementation of this step can construct a more accurate event trigger word library. Through the joint action of multi-modal dynamic fusion and the Transformer-GAT-CRF hybrid network model, event trigger words can be extracted more accurately.

[0062] Step S05: According to the trigger words in the event trigger word library, event extraction is performed in the event corpus to obtain the event types and event arguments corresponding to the trigger words.

[0063] In some embodiments, the trigger words in the event trigger word library are converted into trigger word feature vectors, the trigger word feature vectors are clustered to obtain preliminary event types, and the preliminary event types are aligned to obtain the final event types; the argument words in the event trigger word library are converted into argument word feature vectors, the event argument feature vectors are clustered to obtain preliminary event arguments, and the preliminary event arguments are aligned to obtain the final event arguments.

[0064] Among them, the steps of aligning the preliminary event types to obtain the final event types include: calculating the first cosine similarity between any two event types, setting a first threshold, and determining whether the first cosine similarity is greater than or equal to the first threshold. If so, the two event types are regarded as the same event type; the steps of aligning the preliminary event arguments to obtain the final event arguments include: calculating the second cosine similarity between any two event arguments, setting a second threshold, and determining whether the second cosine similarity is greater than or equal to the second threshold. If so, the two event arguments are regarded as the same event argument.

[0065] Specifically, first, trigger word clustering and preliminary event type determination: The trigger words in the event trigger word library are converted into trigger word feature vectors, and these feature vectors are input into a clustering algorithm (such as the DBSCAN algorithm) for clustering analysis. After clustering, a set of clusters will be obtained, and each cluster represents a class of similar trigger words. The noise points that do not meet the cluster conditions are excluded, the scatter plot is used to visualize the clustering results, and evaluation indicators such as the silhouette coefficient are used to evaluate the clustering quality. Through continuous iterative adjustment, the optimal trigger word clusters are obtained, and the trigger words in the same cluster are classified as the same preliminary event type.

[0066] At the same time, argument word clustering and preliminary event argument determination: The argument words in the event trigger word library are converted into argument word feature vectors, and these feature vectors are input into a clustering algorithm (such as the DBSCAN algorithm) for clustering. After clustering, a set of clusters is obtained, and each cluster represents the event arguments of a certain type of disaster event. Similarly, the noise points are excluded, the clustering results are visualized and the quality is evaluated. After iterative adjustment, the optimal event argument clusters are obtained, and the event arguments in the same cluster are classified as the same preliminary event argument, thereby determining the roles of the event arguments.

[0067] Furthermore, aligning the preliminary event types to obtain the final event types: Calculate the first cosine similarity between any two preliminary event types and set a first threshold. Determine whether the first cosine similarity is greater than or equal to the first threshold. If so, these two event types are regarded as the same event type, thereby obtaining the final event types.

[0068] At the same time, the preliminary event arguments are aligned to obtain the final event arguments: the second cosine similarity between any two preliminary event arguments is calculated, and the second threshold is set. It is determined whether the second cosine similarity is greater than or equal to the second threshold. If the condition is met, it is considered that the two event arguments have a high similarity, and they are regarded as the same or similar event arguments, so that the event arguments are aligned to obtain the final event argument.

[0069] Furthermore, disaster event alignment and event node acquisition: Based on the determined event type, the core trigger word of each event is determined as the event identifier, and the standard structure of the event tuple is defined. The event type is combined with the standardized event argument to form an event tuple set, and the similarity of the two event tuple sets is calculated to find events with similar argument combinations, and similar structural patterns or logical associations are found between different event types to achieve disaster event alignment. Finally, the obtained disaster event tuple is embedded as a vector as a whole, and clustering is performed again using a clustering algorithm. After visualization, quality assessment and iterative adjustment, the optimal event cluster is obtained, and then the nodes of the disaster event can be obtained.

[0070] In this step, the unstructured event corpus is converted into standardized event types and arguments through trigger word and argument feature vectorization, clustering algorithm (such as DBSCAN) and cosine similarity alignment, eliminating semantic redundancy and achieving accurate classification of event elements; at the same time, through the structured expression of event tuples and secondary clustering, the semantic alignment and logical association of disaster event nodes are established, providing highly consistent and parsable disaster event structured data support for disaster chain evolution reasoning, knowledge graph construction and emergency decision-making.

[0071] Step S06: According to the trigger words in the event trigger vocabulary, relationship extraction is performed in the event corpus to obtain semantic relationships, temporal relationships and spatial relationships related to the trigger words, wherein the temporal relationships and spatial relationships are obtained by the improved ELECTRA-BiLSTM model, and the improved ELECTRA-BiLSTM model includes the ELECTRA model, the void convolution layer, the gating mechanism layer, the BiLSTM model and the multi-head attention mechanism layer.

[0072] In some embodiments, the process of acquiring semantic relationships related to trigger words includes: selecting text fragments of multiple trigger words in an event trigger word library to form candidate event pairs; after word segmentation and part-of-speech tagging of the candidate event pairs, constructing a dependency syntax tree to analyze the preliminary semantic relationships between words; based on the preliminary semantic relationships, establishing a mapping rule library from specific to abstract events; performing logical deduction and verification on the mapping rule library to obtain a dependency pattern rule library; and adjusting the trigger word weights in the dependency pattern rule library for optimization to obtain extraction results containing semantic relationships.

[0073] Specifically, firstly, based on the event trigger vocabulary and the event types it identifies, the text corpus of urban rainstorm and flood disaster chain events is scanned to identify and extract text fragments containing more than two trigger words, and these text fragments containing multiple trigger words are constituted into candidate event pairs.

[0074] Further, the preliminary semantic relationship is analyzed: the common associations between disaster events are identified and defined, such as causality, sequence, concurrency, condition, and hyponymy. The candidate event pairs are segmented into words or vocabulary units, and each word is assigned a part-of-speech tag (such as noun, verb, adjective, etc.). Then it is input into the dependency syntactic analysis to determine the dependency relationship between words or vocabulary, and a dependency syntactic tree is constructed based on this. The candidate event pairs are matched, and the specific association relationship between the event relationship and the candidate events is determined and marked, so as to analyze the preliminary semantic relationship between words.

[0075] Furthermore, a mapping rule base is established: a set of abstract events that can summarize the common characteristics of specific urban rainstorm and flood disaster events is defined, and based on the characteristics and attributes of disaster events, a mapping rule base from specific disaster events to abstract disaster events is constructed to map the identified specific disaster events to abstract disaster events.

[0076] Furthermore, the dependency pattern rule base is derived and verified: based on the relationship between the determined specific disaster events, the generalization method is used to infer the possible relationship between abstract events, and then the logical relationship between abstract disaster events is determined by logical reasoning. The accuracy and rationality of the determined logical relationship are verified through expert verification or analogy with historical disaster events, thus obtaining the dependency pattern rule base.

[0077] Furthermore, the extraction results are optimized: in the dependency pattern rule base, an iterative method is used to continuously update and improve the event relationship extraction results by adjusting the trigger word weights, and finally an extraction result containing semantic relationships is obtained.

[0078] In some embodiments, the temporal and spatial relationships related to trigger words are obtained by an improved ELECTRA-BiLSTM model, and the process includes: selecting text fragments of multiple trigger words in the event trigger vocabulary, inputting the pre-trained text encoder ELECTRA model for pre-training, and obtaining corresponding text word vector sequences; inputting the text word vector sequence into the void convolution layer to extract the local-global features of the text; inputting the local-global features into the BiLSTM model to capture the context dependencies; inputting the context dependencies into the multi-head attention mechanism layer to calculate the first attention of the temporal task and the second attention of the spatial task respectively; inputting the first attention and the second attention into the classifier to obtain the classification prediction results of the temporal relationship and the spatial relationship respectively.

[0079] The process of extracting spatial relations is similar to that of extracting temporal relations. First, obtain the text word vector sequence: Select text segments with multiple trigger words from the event trigger word library, use them as input data, and input them into the pre-trained text encoder ELECTRA model for pre-training to obtain the corresponding text word vector sequence.

[0080] Furthermore, extract local-global features: Input the obtained text word vector sequence into a dilated convolutional layer (such as the Dilated Convolutional Neural Network, DGCNN). In this process, parameters such as the number of convolutional layers, the size of the convolutional kernel, and the dilation rate need to be determined, and the dilated convolution formula is calculated to expand the receptive field of the convolutional kernel. Apply an activation function after the convolutional layer to introduce non-linearity, introduce the Gated Recurrent Unit (GRU) to control the information flow, and add residual connections to solve the problem of gradient disappearance in deep networks, and finally extract the local-global features of the text.

[0081] Furthermore, capture context dependencies: Input the extracted local-global features into a BiLSTM model to capture context dependencies and obtain context semantic features.

[0082] Furthermore, calculate attention: Input the context dependencies into the multi-head attention mechanism layer, and calculate the first attention for the temporal task and the second attention for the spatial task respectively. Specifically, linearly transform the output features of the BiLSTM layer to obtain the query matrix, key matrix, and value matrix, split the query matrix into several heads, calculate the attention weights, then use the attention weights to perform weighted summation on the value matrix, merge the outputs of each head and perform a linear transformation to obtain the output of the multi-head attention.

[0083] Finally, input the first attention and the second attention into the classifier respectively, convert them into scores for each category through a linear layer, then calculate the probability of each category, use the cross-entropy loss function to calculate the difference between the predicted probability distribution and the true label, and finally obtain the classification prediction results of the temporal relation and the spatial relation respectively.

[0084] In this step, in terms of semantic relationship acquisition, we start with constructing candidate event pairs, determine preliminary semantic relationships through word segmentation, part-of-speech tagging, and dependency syntactic analysis, then establish a mapping rule base, obtain a dependency pattern rule base through logical deduction and verification, and finally optimize the weights to obtain extraction results containing semantic relationships, thus achieving accurate mining of the semantic relationship of trigger words. For temporal and spatial relationships, we use the improved ELECTRA-BiLSTM model, starting from pre-training to obtain text word vector sequences, extract local-global features through the dilated convolution layer, capture contextual dependencies through the BiLSTM model, and calculate attention through the multi-head attention mechanism layer. Finally, the classification prediction results are obtained by the classifier, which provides strong support for a comprehensive understanding of the complex relationships between events in the urban rainstorm and flood disaster chain, and helps to analyze and respond to disasters more deeply.

[0085] Step S07: Using event types, event arguments, temporal relationships and spatial relationships as event nodes, and semantic relationships as event edges, a disaster chain event graph of rainstorm and flood disasters in the target city is constructed.

[0086] For example, the graph database Neo4j can be used for storage to obtain a disaster chain event map of rainstorm and flood disasters in the target city.

[0087] Example 2

[0088] See also Figure 2 , which is a schematic diagram of the structure of a system for constructing a chain of urban rainstorm and flood disasters according to a second embodiment of the present application, and which includes:

[0089] The data acquisition module 100 is used to acquire research documents and primary historical events and secondary events of rainstorm and flood disasters in the target city, and construct a directed network diagram of the disaster chain;

[0090] An ontology modeling module 200 is used to construct an ontology model of the rainstorm and flood disaster in the target city according to the disaster chain directed network diagram;

[0091] The corpus generation module 300 is used to obtain the second historical event and its secondary events of the rainstorm and flood disaster in the target city for preprocessing, and input the preprocessed second historical event and its secondary events into the ontology model to obtain an event corpus;

[0092] A trigger word extraction module 400 is used to extract event trigger words from the event corpus using a multimodal dynamic fusion model MDF-Net to construct an event trigger word library;

[0093] An event extraction module 500 is used to extract events from the event corpus according to the trigger words in the event trigger word library, and obtain event types and event arguments corresponding to the trigger words;

[0094] A relation extraction module 600 is used to extract relations in the event corpus according to the trigger words in the event trigger lexicon, and obtain semantic relations, temporal relations and spatial relations related to the trigger words, wherein the temporal relations and the spatial relations are obtained by an improved ELECTRA-BiLSTM model, and the improved ELECTRA-BiLSTM model includes an ELECTRA model, a hole convolution layer, a gating mechanism layer, a BiLSTM model and a multi-head attention mechanism layer;

[0095] The graph construction module 700 is used to construct a disaster chain event graph of the rainstorm and flood disaster in the target city with the event type, the event argument, the temporal relationship and the spatial relationship as event nodes and the semantic relationship as the edge of the event.

[0096] A system for constructing a chain of urban rainstorm and flood disasters in an embodiment of the present application may be a device, or a component, integrated circuit, or chip in a terminal. The device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device may be a server, a network attached storage (NAS), a personal computer (PC), etc., which is not specifically limited in the embodiment of the present application.

[0097] In the embodiment of the present application, a system for constructing a chain of urban rainstorm and flood disasters can be a device with an operating system. The operating system can be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0098] The system for constructing a chain diagram of urban rainstorm and flood disasters provided in the embodiment of the present application can achieve Figure 1 In the method embodiment, the various processes of implementing a method for constructing a chain event graph of urban rainstorm and flood disasters are not described here to avoid repetition.

[0099] Optionally, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, each process of the above-mentioned embodiment of a method for constructing a chain cause graph of urban rainstorm and flood disasters is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0100] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned embodiment of a method for constructing a chain of urban rainstorm and flood disasters cause map is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0101] The processor is a processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0102] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0103] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present application.

[0104] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of the present application, those of ordinary skill in the art can also make many forms without departing from the purpose of the present application and the scope protected by the claims, and all of them fall within the protection scope of the present application.

Claims

1. A method for constructing a chain diagram of urban rainstorm and flood disasters, characterized in that: The method comprises: Acquire research literature and the first historical event and its secondary events of rainstorm and flood disasters in the target city, and construct a directed network diagram of the disaster chain, including: acquire research literature on rainstorm and flood disasters in the target city, and organize the research literature into a research review by using a literature research method; acquire the first historical event and its secondary events of rainstorm and flood disasters in the target city, and organize the first historical event and its secondary events into case events by using a case analysis method; construct a disaster chain event set based on the research review and the case events; and draw a directed network diagram of the disaster chain corresponding to the rainstorm and flood disasters in the target city based on the disaster chain event set; According to the directed network diagram of the disaster chain, an ontology model of the rainstorm and flood disaster in the target city is constructed; The second historical event and its secondary events of the rainstorm and flood disaster in the target city are obtained for preprocessing, and the preprocessed second historical event and its secondary events are input into the ontology model to obtain an event corpus; The multimodal dynamic fusion model MDF-Net is used to extract event trigger words from the event corpus to construct an event trigger word library, including: multi-granular annotation of the unstructured text of the event corpus, using the BIOES annotation system and combining the disaster field ontology knowledge, generating fine-grained annotation data through entity boundary enhancement and semantic role annotation; aligning the text data of the event corpus with the real-time meteorological data, geographic information system layer and social media public opinion data of the target city in time and space to generate text alignment data; using the cross-modal attention mechanism of the multimodal dynamic fusion model MDF-Net to extract the association weights of the text keywords and spatiotemporal features of the fine-grained annotation data and the text alignment data to generate a unified multimodal embedding representation; inputting the multimodal embedding representation into a Transformer-GAT-CRF hybrid network model, and the Transformer-GAT-CRF hybrid The pre-training layer of the hybrid network model adopts the ELECTRA model enhanced in the disaster field, and generates context-aware word vectors through masked language modeling and causal prediction tasks; the graph attention layer GAT of the Transformer-GAT-CRF hybrid network model embeds the disaster ontology knowledge graph to capture the topological dependencies between event trigger words, and allocates the weights of events, locations, and time nodes through multi-head attention; the temporal-spatial bidirectional gating layer of the Transformer-GAT-CRF hybrid network model combines BiLSTM and dilated temporal convolution DTCN to model the temporal evolution and spatial diffusion characteristics of events respectively; the contrastive learning module of the Transformer-GAT-CRF hybrid network model optimizes the discrimination of trigger word representation by constructing positive and negative sample pairs; the CRF decoder of the Transformer-GAT-CRF hybrid network model outputs event trigger words; According to the trigger words in the event trigger vocabulary, event extraction is performed in the event corpus to obtain event types and event arguments corresponding to the trigger words, including: converting the trigger words in the event trigger vocabulary into trigger word feature vectors, clustering the trigger word feature vectors to obtain preliminary event types, and aligning the preliminary event types to obtain final event types, specifically including: calculating the first cosine similarity between any two event types, and setting a first threshold, judging whether the first cosine similarity is greater than or equal to the first threshold, and if so, the two event types are regarded as the same event type; converting the argument words in the event trigger vocabulary into argument word feature vectors, clustering the argument word feature vectors to obtain preliminary event arguments, and aligning the preliminary event arguments to obtain final event arguments, specifically including: calculating the second cosine similarity between any two event arguments, and setting a second threshold, judging whether the second cosine similarity is greater than or equal to the second threshold, and if so, the two event arguments are regarded as the same event argument; According to the trigger words in the event trigger word library, relations are extracted in the event corpus to obtain semantic relations, temporal relations and spatial relations related to the trigger words, wherein: The step of obtaining the semantic relationship related to the trigger word includes: selecting multiple text fragments of trigger words in the event trigger word library to form candidate event pairs; after word segmentation and part-of-speech tagging of the candidate event pairs, constructing a dependency syntax tree to analyze the preliminary semantic relationship between words; according to the preliminary semantic relationship, establishing a mapping rule library from specific to abstract events; performing logical deduction and verification on the mapping rule library to obtain a dependency pattern rule library; adjusting the trigger word weight in the dependency pattern rule library for optimization to obtain an extraction result containing semantic relationships; The temporal relationship and the spatial relationship are obtained by an improved ELECTRA-BiLSTM model, which includes an ELECTRA model, a hole convolution layer, a gating mechanism layer, a BiLSTM model and a multi-head attention mechanism layer; including: selecting text fragments of multiple trigger words in the event trigger vocabulary, inputting the pre-trained text encoder ELECTRA model for pre-training, and obtaining corresponding text word vector sequences; inputting the text word vector sequence into the hole convolution layer to extract the local-global features of the text; inputting the local-global features into the BiLSTM model to capture the context dependency; inputting the context dependency into the multi-head attention mechanism layer to calculate the first attention of the temporal task and the second attention of the spatial task respectively; obtaining the classification prediction results of the temporal relationship and the spatial relationship according to the first attention and the second attention respectively; A disaster chain event graph of rainstorm and flood disasters in the target city is constructed by taking the event type, the event argument, the temporal relationship and the spatial relationship as event nodes and taking the semantic relationship as the edge of the event.

2. A system for constructing a chain diagram of urban rainstorm and flood disasters, characterized in that: The system comprises: The data acquisition module is used to acquire research literature and the first historical event and its secondary events of rainstorm and flood disasters in the target city, and construct a directed network diagram of the disaster chain, including: acquiring research literature on rainstorm and flood disasters in the target city, and using the literature research method to organize the research literature into a research review; acquiring the first historical event and its secondary events of rainstorm and flood disasters in the target city, and using the case analysis method to organize the first historical event and its secondary events into case events; constructing a disaster chain event set based on the research review and the case events; and drawing a directed network diagram of the disaster chain corresponding to the rainstorm and flood disasters in the target city based on the disaster chain event set; An ontology modeling module, used to construct an ontology model of the rainstorm and flood disaster in the target city according to the directed network diagram of the disaster chain; A corpus generation module is used to obtain the second historical event and its secondary events of the rainstorm and flood disaster in the target city for preprocessing, and input the preprocessed second historical event and its secondary events into the ontology model to obtain an event corpus; The trigger word extraction module is used to extract event trigger words from the event corpus using the multimodal dynamic fusion model MDF-Net to construct an event trigger word library, including: multi-granular annotation of the unstructured text of the event corpus, using the BIOES annotation system and combining the disaster field ontology knowledge, generating fine-grained annotation data through entity boundary enhancement and semantic role annotation; aligning the text data of the event corpus with the real-time meteorological data, geographic information system layer and social media public opinion data of the target city in time and space to generate text alignment data; using the cross-modal attention mechanism of the multimodal dynamic fusion model MDF-Net to extract the association weights of the text keywords and spatiotemporal features of the fine-grained annotation data and the text alignment data, and generating a unified multimodal embedding representation; inputting the multimodal embedding representation into the Transformer-GAT-CRF hybrid network model, the Transformer-GAT -The pre-training layer of the CRF hybrid network model adopts the ELECTRA model enhanced in the disaster field, and generates context-aware word vectors through masked language modeling and causal prediction tasks; the graph attention layer GAT of the Transformer-GAT-CRF hybrid network model embeds the disaster ontology knowledge graph to capture the topological dependencies between event trigger words, and allocates the weights of events, locations, and time nodes through multi-head attention; the temporal-spatial bidirectional gating layer of the Transformer-GAT-CRF hybrid network model combines BiLSTM and dilated temporal convolution DTCN to model the temporal evolution and spatial diffusion characteristics of events respectively; the contrastive learning module of the Transformer-GAT-CRF hybrid network model optimizes the discrimination of trigger word representation by constructing positive and negative sample pairs; the CRF decoder of the Transformer-GAT-CRF hybrid network model outputs event trigger words; An event extraction module is used to extract events in the event corpus according to the trigger words in the event trigger vocabulary, and obtain the event type and event argument corresponding to the trigger words, including: converting the trigger words in the event trigger vocabulary into trigger word feature vectors, clustering the trigger word feature vectors to obtain preliminary event types, and aligning the preliminary event types to obtain final event types, specifically including: calculating the first cosine similarity between any two event types, and setting a first threshold, judging whether the first cosine similarity is greater than or equal to the first threshold, and if so, the two event types are regarded as the same event type; converting the argument words in the event trigger vocabulary into argument word feature vectors, clustering the argument word feature vectors to obtain preliminary event arguments, and aligning the preliminary event arguments to obtain final event arguments, specifically including: calculating the second cosine similarity between any two event arguments, and setting a second threshold, judging whether the second cosine similarity is greater than or equal to the second threshold, and if so, the two event arguments are regarded as the same event argument; The relationship extraction module is used to extract relationships in the event corpus according to the trigger words in the event trigger word library, and obtain the semantic relationship, temporal relationship and spatial relationship related to the trigger words, wherein: The step of obtaining the semantic relationship related to the trigger word includes: selecting multiple text fragments of trigger words in the event trigger word library to form candidate event pairs; after word segmentation and part-of-speech tagging of the candidate event pairs, constructing a dependency syntax tree to analyze the preliminary semantic relationship between words; according to the preliminary semantic relationship, establishing a mapping rule library from specific to abstract events; performing logical deduction and verification on the mapping rule library to obtain a dependency pattern rule library; adjusting the trigger word weight in the dependency pattern rule library for optimization to obtain an extraction result containing semantic relationships; The temporal relationship and the spatial relationship are obtained by an improved ELECTRA-BiLSTM model, which includes an ELECTRA model, a hole convolution layer, a gating mechanism layer, a BiLSTM model and a multi-head attention mechanism layer; including: selecting text fragments of multiple trigger words in the event trigger vocabulary, inputting the pre-trained text encoder ELECTRA model for pre-training, and obtaining corresponding text word vector sequences; inputting the text word vector sequence into the hole convolution layer to extract the local-global features of the text; inputting the local-global features into the BiLSTM model to capture the context dependency; inputting the context dependency into the multi-head attention mechanism layer to calculate the first attention of the temporal task and the second attention of the spatial task respectively; obtaining the classification prediction results of the temporal relationship and the spatial relationship according to the first attention and the second attention respectively; The graph construction module is used to construct a disaster chain event graph of the rainstorm and flood disaster in the target city with the event type, the event argument, the temporal relationship and the spatial relationship as event nodes and the semantic relationship as the edge of the event.

3. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of a method for constructing a chain cause graph of urban rainstorm and flood disasters as described in any one of claims 1 to 2 are implemented.

4. A readable storage medium, characterized in that: The readable storage medium stores programs or instructions, and when the programs or instructions are executed by the processor, the steps of the method for constructing a chain cause graph of urban rainstorm and flood disasters as described in any one of claims 1-2 are implemented.

Images