Natural disaster event extraction method and system
By performing multi-grained sharding and fusion of sentence vectors and perceived and fusion processing of event vectors, an event prediction model is constructed, which solves the problem that multiple event overlapping sentences cannot be handled efficiently and accurately in the existing technology, and improves the efficiency and accuracy of extraction of natural disaster events.
Patent Information
- Application Number
- CN202510050094.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-13
- Publication Date
- 2025-05-06
AI Technical Summary
Existing event extraction methods cannot efficiently and accurately process sentences with overlapping multiple events, resulting in low processing efficiency and accuracy.
By performing multi-grained fragmentation and fusion processing on the sentence vector, a multi-grained sentence representation is obtained, and the event vector is perceived and fusion processing is performed to obtain event-aware sentence representation. Then, based on these representations, the event prediction model is built to output the key event information corresponding to the disaster data.
It realizes efficient and accurate extraction of overlapping sentences of multiple events, and improves the efficiency and accuracy of natural disaster incident handling.
Smart Images
Figure CN119940354A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of event extraction, and in particular relates to a method and system for extracting natural disaster events. Background Art
[0002] Natural disasters refer to natural phenomena that endanger human survival or damage the human living environment. In order to minimize the property losses caused by natural disasters, it is usually necessary to conduct in-depth analysis and summary of related disaster events. Before conducting in-depth analysis and summary of related disaster events, it is necessary to collect and summarize the disaster events. The traditional method of collecting and summarizing disaster events is manual, which has the disadvantages of low processing efficiency, easy mistakes by workers, and low accuracy.
[0003] With the development of artificial intelligence technology, event extraction is an important task in the field of natural language processing (NLP). It aims to identify specific types of events from natural language texts and extract key information related to these events, such as event trigger words, (arguments), time, location, etc.
[0004] Therefore, statistically summarizing data related to natural disasters through event extraction can improve the efficiency of natural disaster event processing. When extracting events from data related to natural disasters, most existing methods assume that only one event appears in a sentence. In real scenarios, a sentence related to natural disasters may contain multiple overlapping events. Existing event extraction methods cannot efficiently and accurately extract sentences with multiple overlapping events. Summary of the invention
[0005] The purpose of the present invention is to provide a method and system for extracting natural disaster events, so as to solve the problem that the existing event extraction method cannot efficiently and accurately extract sentences with multiple overlapping events.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for extracting natural disaster events, the method comprising: Obtain sample data of disaster events, pre-process the sample data of disaster events, and obtain initial sentences; Encode and sample the initial sentence to obtain the sentence vector and event vector; Perform multi-granularity segmentation and fusion processing on sentence vectors to obtain multi-granularity sentence representation; The event vectors are perceived and fused to obtain event-aware sentence representations; Based on multi-granularity sentence representation and event-aware sentence representation, an event prediction model is constructed. The event prediction model is used to take disaster data as input and output key event information corresponding to the disaster data.
[0007] Preferably, the method further comprises: Obtain disaster data; Input disaster data into the event prediction model to obtain key event information.
[0008] Preferably, an event prediction model is constructed based on multi-granularity sentence representation and event-aware sentence representation, including: The multi-granular sentence representation and event-aware sentence representation are fused to obtain the final sentence representation; Calculating a fragment-aware score and an object-aware score for each word pair in the final representation of the sentence; the fragment-aware score is used to characterize the probability that the word pair is a fragment, and the object-aware score is used to characterize the probability that the word pair is an object, wherein a fragment includes: an argument and / or a trigger word; An event prediction model is constructed based on the fragment-aware score and object-aware score of each word pair in the final representation of the sentence.
[0009] Preferably, the sentence vector is subjected to multi-granularity segmentation and fusion processing to obtain a multi-granularity sentence representation, including: Perform segmentation feature extraction on the sentence vector to obtain the long segment features and short segment features of the sentence; The long segment features and short segment features of the sentence are concatenated to obtain the sentence features; The sentence vector is fused with the sentence features to obtain a multi-granularity sentence representation.
[0010] Preferably, the sentence vector is subjected to segmentation feature extraction to obtain the long segment features and short segment features of the sentence, including: The sentence vector is processed based on the attention mechanism to obtain a first feature vector, wherein the first feature vector includes a plurality of continuous feature segments, and the first feature vector is used as a long segment feature of the sentence; Performing multi-stage convolution on the first feature vector to obtain a sentence fragment representation, wherein the number of feature fragments in the sentence fragment representation is less than the number of feature fragments in the first feature vector; Multi-stage convolution is performed on the sentence fragment representation to obtain the short fragment features of the sentence.
[0011] Preferably, a multi-stage convolution is performed on the first feature vector to obtain a sentence fragment representation, including: Using the first feature vector as a first short segment division vector, processing the first short segment division vector based on an attention mechanism to obtain a processed first short segment division vector; Performing convolution processing on the processed first short segment division vector with a preset convolution kernel and a preset step to obtain a first vector, wherein the first vector includes a plurality of continuous feature segments, and the number of feature segments in the first vector is less than the number of feature segments in the first feature vector; The first vector is used as a new first short segment division vector, and the first short segment division vector is repeatedly processed based on the attention mechanism to obtain the processed first short segment division vector, and the processed first short segment division vector is convolved with a preset convolution kernel and a preset step to obtain the first vector; until the number of feature segments in the first vector reaches a preset number, then the sentence short segment representation is obtained.
[0012] Preferably, a multi-stage convolution is performed on the sentence fragment representation to obtain the short fragment features of the sentence, including: The sentence fragment representation is used as the second short fragment segmentation vector, and the second short fragment segmentation vector is processed based on the attention mechanism to obtain a processed second short fragment segmentation vector; Performing convolution processing on the processed second short segment division vector with a preset convolution kernel and a preset step to obtain a second vector, wherein the second vector includes a plurality of continuous feature segments, and the number of feature segments in the second vector is greater than the number of feature segments in the second short segment division vector; The second vector is used as a new second short segment division vector, and the second short segment division vector is repeatedly processed based on the attention mechanism to obtain a processed second short segment division vector, and the processed second short segment division vector is convolved with a preset convolution kernel and a preset step to obtain a second vector; until the number of feature segments in the second vector is equal to the number of feature segments in the first feature vector, at which time the short segment features of the sentence are obtained.
[0013] Preferably, the preset convolution kernel of the convolution processing is 2, and the preset step of the convolution processing is 1.
[0014] Preferably, the event vector is sensed and fused to obtain an event-aware sentence representation, including: Extract features from event vectors to obtain global event features; The event vector is fused with the global event feature to obtain event-aware sentence representation.
[0015] In a second aspect, the present invention provides a natural disaster event extraction system, the system is used to implement the above-mentioned natural disaster event extraction method, the system comprises: A data processing module is used to obtain sample data of disaster events, pre-process the sample data of disaster events, and obtain initial sentences; The encoding and sampling module is used to encode and sample the initial sentence to obtain the sentence vector and event vector; The first fusion module is used to perform multi-granularity segmentation and fusion processing on the sentence vector to obtain multi-granularity sentence representation; The second fusion module is used to perceive and fuse event vectors to obtain event-aware sentence representation; The model building module is used to build an event prediction model based on multi-granularity sentence representation and event-aware sentence representation. The event prediction model is used to take disaster data as input and output key event information corresponding to the disaster data.
[0016] Beneficial effects: The present invention can obtain a multi-granularity sentence representation by performing multi-granularity segmentation and fusion on sentence vectors. The multi-granularity sentence representation has the characteristics of different language granularities, that is, it can effectively extract local features in the text and capture longer-range dependencies and context information in the sentence; then, the event vector obtained by sampling the initial sentence is sensed and fused to obtain an event-aware sentence representation; then, based on the multi-granularity sentence representation and the event-aware sentence representation, an event prediction model is constructed. The event prediction model is used to characterize the mapping relationship between words and events in a sentence, which can solve the problem of multiple event overlaps and improve the extraction efficiency and accuracy of multiple event overlapping sentences. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings: Figure 1 is a flow chart of a method for extracting natural disaster events provided by one embodiment of the present invention; Figure 2 It is a block diagram of a natural disaster event extraction system provided by one embodiment of the present invention. DETAILED DESCRIPTION
[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.
[0019] Embodiment 1 Figure 1is a flow chart of a method for extracting natural disaster events provided by one embodiment of the present invention, such as Figure 1 As shown, this embodiment provides a method for extracting natural disaster events, the method comprising: Step S10: Obtain sample data of disaster events, pre-process the sample data of disaster events, and obtain initial sentences.
[0020] In this embodiment, the preprocessing steps include data cleaning, format conversion, extraction of key information, etc., among which data cleaning mainly includes: deleting duplicate data, filtering irrelevant data, removing outliers and data consistency, etc.; format conversion mainly includes: data type conversion, data structure conversion, encoding conversion, file format conversion and date format conversion, etc.
[0021] Step S20: Encode and sample the initial sentence to obtain a sentence vector and an event vector.
[0022] In this embodiment, the BERT (Bidirectional Encoder Representations from Transformers) model can be used to encode the initial sentence to obtain a sentence vector. The sentence vector of this implementation has multiple words. After each word is encoded by BERT to encode the initial sentence, the maximum pooling is used to generate an embedded representation of each word. In this embodiment, the event type corresponding to the initial sentence is sampled using positive and negative sampling strategies to obtain an event vector.
[0023] Step S30: performing multi-granularity segmentation and fusion processing on the sentence vector to obtain a multi-granularity sentence representation; In this embodiment, although the convolution operation can effectively extract local features in the text, it also has the problem of not being able to fully capture the longer-range dependencies and contextual information in the sentence. Therefore, the sentence vector is segmented and fused at multiple granularities to obtain a multi-granularity sentence representation, including: Step S301: extracting fragment features from the sentence vector to obtain long fragment features and short fragment features of the sentence. The steps of extracting fragment features from the sentence vector in this embodiment are as follows: Step a10: Process the sentence vector based on the attention mechanism to obtain a first feature vector, which contains multiple continuous feature segments, and uses the first feature vector as the long segment feature of the sentence.
[0024] A feature segment in the first feature vector of this embodiment represents a word in an initial sentence, and the number of words in an initial sentence is the same as the number of feature segments corresponding to the initial sentence.
[0025] Step a20: performing multi-stage convolution on the first feature vector to obtain a sentence fragment representation, wherein the number of feature segments in the sentence fragment representation is less than the number of feature segments in the first feature vector.
[0026] In this embodiment, the number of convolution stages is determined by the number of feature segments in the first feature vector and the number of feature segments in the sentence fragment representation. For example, if the number of feature segments in the first feature vector is 5 and the number of feature segments in the sentence fragment representation is 1, there are a total of 5 stages of convolution operations. Each convolution operation uses a filter with a convolution kernel of 2 and a convolution step of 1 for convolution operation. After each stage of convolution operation, the number of feature segments obtained in the next stage will be reduced. Multiple stages of convolution operations are repeated until there is only one feature segment in the fifth stage, at which time the final sentence fragment representation can be obtained.
[0027] Therefore, the specific steps of performing multi-stage convolution on the first feature vector to obtain the sentence fragment representation are: Step a201: using the first feature vector as the first short segment division vector, processing the first short segment division vector based on the attention mechanism to obtain a processed first short segment division vector; Step a202: performing convolution processing on the processed first short segment division vector with a preset convolution kernel and a preset step to obtain a first vector, wherein the first vector contains a plurality of continuous feature segments, and the number of feature segments in the first vector is less than the number of feature segments in the first feature vector; Step a203: Use the first vector as a new first short segment division vector, repeatedly perform processing on the first short segment division vector based on the attention mechanism, obtain the processed first short segment division vector, perform convolution processing on the processed first short segment division vector with a preset convolution kernel and a preset step, and obtain the first vector; until the number of feature segments in the first vector reaches a preset number, then a sentence segment representation is obtained.
[0028] Step a30: Perform multi-stage convolution on the sentence fragment representation to obtain the short fragment features of the sentence.
[0029] In this embodiment, although the convolution operation can effectively extract local features in the text, it also has the problem of not being able to fully capture the longer-range dependencies and contextual information in the sentence; that is, the text with a relatively fine word granularity represented by the sentence fragment is relatively fine, and cannot directly capture the semantic information of the context. In order to enable the established model to simultaneously focus on the long-range dependencies of the input sequence and learn contextual information; therefore, multi-stage convolution is performed on the sentence fragment representation to obtain the short segment features of the sentence. Specifically, the multi-stage convolution of the sentence fragment representation includes the following steps: Step a301: Perform multi-stage convolution on the sentence fragment representation to obtain the short fragment features of the sentence, including: Step a302: taking the sentence segment representation as the second short segment segmentation vector, and processing the second short segment segmentation vector based on the attention mechanism to obtain a processed second short segment segmentation vector; Step a303: Convolution processing is performed on the processed second short segment division vector with a preset convolution kernel and a preset step to obtain a second vector, wherein the second vector contains multiple continuous feature segments, and the number of feature segments in the second vector is greater than the number of feature segments in the second short segment division vector; wherein the preset convolution kernel of the convolution processing is 2, and the preset step of the convolution processing is 1.
[0030] Step a304: Use the second vector as a new second short segment division vector, repeatedly perform processing on the second short segment division vector based on the attention mechanism, obtain the processed second short segment division vector, perform convolution processing on the processed second short segment division vector with a preset convolution kernel and a preset step, and obtain the second vector; until the number of feature segments in the second vector is equal to the number of feature segments in the first feature vector, at which time the short segment features of the sentence are obtained.
[0031] In this embodiment, in step a303, the number of feature segments in the second short segment division vector before the convolution operation is less than the number of feature segments in the second vector after the convolution operation. Therefore, before the convolution operation, when the sentence segment representation is used as the second short segment division vector, a supplementary segment is added before the first feature segment and after the last feature segment represented by the sentence segment (the supplementary segment may be a zero vector), and when the second vector is used as a new second short segment division vector, a supplementary segment is added before the first feature segment and after the last feature segment of the second vector (the supplementary segment may be a zero vector).
[0032] Step S302: concatenate the long segment features and the short segment features of the sentence to obtain the sentence features; In this embodiment, the short fragment features of the sentence include fragments of different granularities. After the long fragment features and the short fragment features of the sentence are spliced, the discrete information of the fragments of different granularities in the short fragment features of the sentence is transferred to the long fragment features of the sentence, and the long fragment features of the sentence can provide contextual information complementary to the short fragment features; therefore, the sentence features obtained after splicing include more comprehensive and accurate contextual semantic information.
[0033] Step S303: Fuse the sentence vector with the sentence feature to obtain a multi-granularity sentence representation.
[0034] In this embodiment, the gating mechanism can be used to fuse the sentence vector and sentence features. The gating mechanism is an important technology in deep learning models. The gating mechanism allows the network to selectively transmit information, thereby better controlling the gradient flow and data flow, solving the long-term dependency problem, and improving learning efficiency. The gating mechanism is used to fuse the sentence vector and sentence features and filter out unnecessary information.
[0035] Since sentence features contain more comprehensive and accurate contextual semantic information, after the complete fusion of sentence vectors and sentence features, the multi-granularity sentence representation has features of different language granularities, which can effectively extract local features in the text and capture longer-range dependencies and contextual information in the sentence.
[0036] Step S40: Perform perception and fusion processing on the event vector to obtain event-aware sentence representation.
[0037] The task goal of this embodiment is to establish the relationship between words of event types, so it is necessary to perceive and fuse the event vector. Each sentence in the sample data of this embodiment is annotated with the corresponding natural disaster event type. The event type corresponding to the initial sentence is sampled using positive and negative sampling strategies to obtain the event vector, and then the attention mechanism is used to model the interaction between events and between events and texts (sentence vectors). The global event embedding representation of each initial sentence can be obtained, and the global event embedding representation of all sentences is used as the global event feature; finally, the global event embedding representation is fused with the event vector using the gating mechanism to obtain the event-aware sentence representation.
[0038] Specifically, the steps of perceiving and fusing event vectors are as follows: Step S401: extracting features from event vectors to obtain global event features.
[0039] Step S402: Fuse the event vector with the global event feature to obtain event-aware sentence representation.
[0040] Step S50: constructing an event prediction model based on the multi-granularity sentence representation and the event-aware sentence representation, wherein the event prediction model is used to take the disaster data as input and output key event information corresponding to the disaster data.
[0041] Specifically, based on multi-granular sentence representation and event-aware sentence representation, an event prediction model is constructed, including: Step S501: Fuse the multi-granularity sentence representation and the event-aware sentence representation to obtain the final sentence representation.
[0042] In this embodiment, a gating mechanism is also used to fuse multi-granularity sentence representation and event-aware sentence representation; in actual prediction tasks, the gating mechanism dynamically selects input information according to different tasks or context conditions, improves the generalization ability of the model in different situations, and enables the model to better adapt to different data distributions.
[0043] Step S502: Calculate the segment perception score and object perception score of each word pair in the final representation of the sentence; the segment perception score is used to characterize the probability that the word pair is a segment, and the higher the probability that the word pair is a segment, the larger the segment perception score; the object perception score is used to characterize the probability that the word pair is an object, and the higher the probability that the word pair is an object, the larger the object perception score, wherein the segment includes: arguments and / or trigger words; Step S503: construct an event prediction model based on the segment perception score and object perception score of each word pair in the final representation of the sentence.
[0044] In this embodiment, the event extraction task is transformed into a word relationship detection task of the target event, which aims to identify each word pair ( x i , x j ) of the fragment relation S and object relation R. Among them, the fragment relation S represents the start and end words of the trigger word fragment or argument fragment, and the object relation R represents the x i The argument contains x j The event triggered by the trigger word plays a certain argument role, denoted as R-*, where * represents the argument role; in this embodiment, the event prediction model adopts a support vector machine, takes multi-granularity sentence representation and event-aware sentence representation as the input of the support vector machine, takes the inter-word relationship of the target event as the label, trains the support vector machine, and uses the trained support vector machine as the event prediction model; the loss function of the event prediction model is constructed by using the fragment perception score and object perception score of each word pair in the final representation of the sentence.
[0045] That is, in order to better jointly predict the fragment relationship and role relationship between word pairs, the word pairs are fused with the distance information, and the fragment-aware score and object-aware score of the embedded representation of each word pair are calculated separately; then the fragment-aware score and the object-aware score are combined. The training goal is to establish the loss function of the event prediction model with the combined fragment-aware score and object-aware score, and to minimize the total loss of the loss function.
[0046] For each event type, the trigger word segment or argument segment of the target event type has the beginning and end indexes, and the trigger word or argument segment is decoded to effectively solve the trigger word overlap problem (a word is used as a trigger word in multiple events) and the argument overlap problem (a word plays different roles as an argument in multiple events). In addition, the obtained role relationship is used to further match trigger words and arguments, which can effectively solve the argument overlap problem (a word plays different roles in an event as an argument).
[0047] As a further optimization of this embodiment, the method further includes: Step b10: Obtain disaster data; Step b20: Input the disaster data into the event prediction model to obtain key event information.
[0048] The present invention can obtain a multi-granularity sentence representation by performing multi-granularity segmentation and fusion on sentence vectors. The multi-granularity sentence representation has the characteristics of different language granularities, that is, it can effectively extract local features in the text and capture longer-range dependencies and context information in the sentence; then, the event vector obtained by sampling the initial sentence is sensed and fused to obtain an event-aware sentence representation; then, based on the multi-granularity sentence representation and the event-aware sentence representation, an event prediction model is constructed. The event prediction model is used to characterize the mapping relationship between words and events in a sentence, which can solve the problem of multiple event overlaps and improve the extraction efficiency and accuracy of multiple event overlapping sentences.
[0049] Embodiment 2 Figure 2 is a block diagram of a natural disaster event extraction system provided by an embodiment of the present invention, such as Figure 2 As shown, this embodiment provides a natural disaster event extraction system, the system is used to implement the natural disaster event extraction method of embodiment 1, and the system includes: A data processing module is used to obtain sample data of disaster events, pre-process the sample data of disaster events, and obtain initial sentences; The encoding and sampling module is used to encode and sample the initial sentence to obtain the sentence vector and event vector; The first fusion module is used to perform multi-granularity segmentation and fusion processing on the sentence vector to obtain multi-granularity sentence representation; The second fusion module is used to perceive and fuse event vectors to obtain event-aware sentence representation; The model building module is used to build an event prediction model based on multi-granularity sentence representation and event-aware sentence representation. The event prediction model is used to take disaster data as input and output key event information corresponding to the disaster data.
[0050] The present invention can obtain a multi-granularity sentence representation by performing multi-granularity segmentation and fusion on sentence vectors. The multi-granularity sentence representation has the characteristics of different language granularities, that is, it can effectively extract local features in the text and capture longer-range dependencies and context information in the sentence; then, the event vector obtained by sampling the initial sentence is sensed and fused to obtain an event-aware sentence representation; then, based on the multi-granularity sentence representation and the event-aware sentence representation, an event prediction model is constructed. The event prediction model is used to characterize the mapping relationship between words and events in a sentence, which can solve the problem of multiple event overlaps and improve the extraction efficiency and accuracy of multiple event overlapping sentences.
[0051] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0052] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0053] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.
Claims
1. A method for extracting natural disaster events, characterized in that: The method comprises: Obtain sample data of disaster events, pre-process the sample data of disaster events, and obtain initial sentences; Encode and sample the initial sentence to obtain the sentence vector and event vector; Perform multi-granularity segmentation and fusion processing on sentence vectors to obtain multi-granularity sentence representation; The event vectors are perceived and fused to obtain event-aware sentence representations; Based on multi-granularity sentence representation and event-aware sentence representation, an event prediction model is constructed. The event prediction model is used to take disaster data as input and output key event information corresponding to the disaster data.
2. The method for extracting natural disaster events according to claim 1, characterized in that: The method further comprises: Obtain disaster data; Input disaster data into the event prediction model to obtain key event information.
3. The method for extracting natural disaster events according to claim 1, characterized in that: Based on multi-granular sentence representation and event-aware sentence representation, an event prediction model is constructed, including: The multi-granular sentence representation and event-aware sentence representation are fused to obtain the final sentence representation; Calculating a fragment-aware score and an object-aware score for each word pair in the final representation of the sentence; the fragment-aware score is used to characterize the probability that the word pair is a fragment, and the object-aware score is used to characterize the probability that the word pair is an object, wherein a fragment includes: an argument and / or a trigger word; An event prediction model is constructed based on the fragment-aware score and object-aware score of each word pair in the final representation of the sentence.
4. The method for extracting natural disaster events according to claim 1, characterized in that: The sentence vectors are segmented and fused at multiple granularities to obtain multi-granularity sentence representations, including: Perform segmentation feature extraction on the sentence vector to obtain the long segment features and short segment features of the sentence; The long segment features and short segment features of the sentence are concatenated to obtain the sentence features; The sentence vector is fused with the sentence features to obtain a multi-granularity sentence representation.
5. The method for extracting natural disaster events according to claim 4, characterized in that: Perform segmentation feature extraction on the sentence vector to obtain the long segment features and short segment features of the sentence, including: The sentence vector is processed based on the attention mechanism to obtain a first feature vector, wherein the first feature vector includes a plurality of continuous feature segments, and the first feature vector is used as a long segment feature of the sentence; Performing multi-stage convolution on the first feature vector to obtain a sentence fragment representation, wherein the number of feature fragments in the sentence fragment representation is less than the number of feature fragments in the first feature vector; Multi-stage convolution is performed on the sentence fragment representation to obtain the short fragment features of the sentence.
6. The method for extracting natural disaster events according to claim 5, characterized in that: Perform multi-stage convolution on the first feature vector to obtain a sentence fragment representation, including: Using the first feature vector as a first short segment division vector, processing the first short segment division vector based on an attention mechanism to obtain a processed first short segment division vector; Performing convolution processing on the processed first short segment division vector with a preset convolution kernel and a preset step to obtain a first vector, wherein the first vector includes a plurality of continuous feature segments, and the number of feature segments in the first vector is less than the number of feature segments in the first feature vector; The first vector is used as a new first short segment division vector, and the first short segment division vector is repeatedly processed based on the attention mechanism to obtain the processed first short segment division vector, and the processed first short segment division vector is convolved with a preset convolution kernel and a preset step to obtain the first vector; until the number of feature segments in the first vector reaches a preset number, a sentence short segment representation is obtained.
7. The method for extracting natural disaster events according to claim 6, characterized in that: Perform multi-stage convolution on the sentence fragment representation to obtain the short fragment features of the sentence, including: The sentence fragment representation is used as the second short fragment segmentation vector, and the second short fragment segmentation vector is processed based on the attention mechanism to obtain a processed second short fragment segmentation vector; Performing convolution processing on the processed second short segment division vector with a preset convolution kernel and a preset step to obtain a second vector, wherein the second vector includes a plurality of continuous feature segments, and the number of feature segments in the second vector is greater than the number of feature segments in the second short segment division vector; The second vector is used as a new second short segment division vector, and the second short segment division vector is repeatedly processed based on the attention mechanism to obtain a processed second short segment division vector, and the processed second short segment division vector is convolved with a preset convolution kernel and a preset step to obtain a second vector; until the number of feature segments in the second vector is equal to the number of feature segments in the first feature vector, at which time the short segment features of the sentence are obtained.
8. The method for extracting natural disaster events according to claim 6 or 7, characterized in that: The preset convolution kernel of the convolution process is 2, and the preset step of the convolution process is 1.
9. The method for extracting natural disaster events according to claim 6, characterized in that: The event vectors are perceived and fused to obtain event-aware sentence representations, including: Extract features from event vectors to obtain global event features; The event vector is fused with the global event feature to obtain event-aware sentence representation.
10. A natural disaster event extraction system, the system is used to implement the natural disaster event extraction method according to any one of claims 1 to 9, characterized in that: The system comprises: A data processing module is used to obtain sample data of disaster events, pre-process the sample data of disaster events, and obtain initial sentences; The encoding and sampling module is used to encode and sample the initial sentence to obtain the sentence vector and event vector; The first fusion module is used to perform multi-granularity segmentation and fusion processing on the sentence vector to obtain multi-granularity sentence representation; The second fusion module is used to perceive and fuse event vectors to obtain event-aware sentence representation; The model building module is used to build an event prediction model based on multi-granularity sentence representation and event-aware sentence representation. The event prediction model is used to take disaster data as input and output key event information corresponding to the disaster data.