Typhoon disaster situation position extraction and classification method and system based on social media

Abstract

The invention discloses a typhoon disaster situation position extraction and classification method and system based on social media, and the method comprises the steps: collecting text data, and carrying out the preprocessing of the text data, and generating a text vector; multi-hidden-layer semantic representations of text vectors are extracted through a BERT model and fused, an auxiliary classifier is used for receiving output of all hidden layers, and intermediate results of position extraction and disaster situation classification are generated to accelerate training; identifying a place name entity through a BiLSTM-CRF model and outputting a BIO tag sequence to complete position extraction; the weight of a place name entity in the text is calculated through a place name enhanced weight mechanism, semantic features related to the position in the disaster situation classification task are enhanced, multi-label classification is carried out through a CNN model, and a disaster situation classification result is output; performing joint optimization training on the two task models through a multi-task loss function; and inputting a text for typhoon disaster situation position extraction and classification. According to the method, disaster situation position information can be extracted from social media in a high-precision mode, multi-label disaster situation classification is achieved, and comprehensive spatial-temporal feature analysis is provided for typhoon disaster assessment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of text data processing, and particularly relates to a typhoon disaster situation location extraction and classification method and system based on social media. BACKGROUND

[0002] Mining location and disaster information from social media needs to overcome the problem of deep learning network for content understanding such as text, image and video. Taking text as an example, text representation plays a crucial role in converting text into vectors containing semantic information, and constitutes the basis of various complex text tasks. Scholars have extensively studied text representation and developed relatively mature technologies, including bag-of-words-based models and Transformer-based models. Among them, BERT (Bidirectional Encoder Representations from Transformers) is pre-trained on large datasets to capture the bidirectional context of text and performs well in various downstream tasks, so it is widely used. The output of different hidden layers of BERT captures different language nuances and performs differently in different tasks. However, most studies only rely on the output of the last layer of BERT, ignoring the semantic information of the hidden layer, resulting in insufficient comprehensiveness of text representation.

[0003] In terms of location extraction, typical location information of social media text includes user registration location and geotagged location. User registration location is established when the user registers on the social media platform, and the number of geotagged text in social media text is small, so neither of the two location information can accurately and comprehensively reflect the actual location of social media text. In order to solve this problem, the method based on place name dictionary extracts location by matching characters with place name dictionary. Although this method is simple and direct to use, it needs to spend a long time to manually create rule templates and place name dictionaries. The place name entity recognition method based on deep learning, including CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) based models, significantly improves the efficiency and accuracy of location information extraction, but its performance still needs to be further improved.

[0004] In terms of disaster classification, the classical methods include rule-based methods and machine learning-based methods. Rule-based methods rely on pre-defined rules to classify text, but depend on domain expertise. Machine learning-based methods learn the relationship between text and labels through hand-crafted features. These methods usually use single-label classification, limiting their ability to capture complete semantic information in the text. For example, "During the typhoon, the road was severely flooded and many tunnels were impassable" expresses the impact of the typhoon on waterlogging and traffic, but single-label classification methods can only identify one of the typhoon disaster categories. Deep learning has driven the development of text classification models, making multi-label classification feasible. For example, models based on CNN, RNN and attention mechanisms have greatly improved text classification accuracy and are widely used in disaster situation awareness. However, current methods usually treat location extraction and category classification as independent tasks, failing to effectively combine the two for collaborative analysis, limiting the comprehensiveness and accuracy of typhoon disaster spatio-temporal feature evaluation. SUMMARY

[0005] The purpose of the present application is to overcome the problems in the prior art and provide a multi-task learning framework that can simultaneously extract location information and classify disaster categories from social media, to meet the complex needs of typhoon disaster emergency rescue under big data, and to achieve accurate and comprehensive typhoon disaster evaluation.

[0006] To achieve the above-mentioned purpose, the first aspect of the present application provides a typhoon disaster location extraction and classification method based on social media, comprising the following steps: S1, collecting typhoon-related social media text data, and preprocessing the text to generate text vectors and dividing them into a training set; S2, extracting the multi-hidden layer semantic representation of the text vector through the BERT model, and fusing it to generate a comprehensive text representation vector, using an auxiliary classifier to receive the output of each hidden layer to generate intermediate results for location extraction and disaster classification to speed up subsequent model training; S3, simultaneously performing the location extraction task and the disaster classification task based on a multi-task learning framework; The location extraction task identifies geographical entity and outputs BIO label sequence through BiLSTM-CRF model, completing location extraction; The disaster classification task dynamically calculates the weight of geographical entity in the text through a geographical name enhanced weight mechanism, strengthens the semantic features related to location in the disaster classification task, and performs multi-label classification through a CNN model to output the disaster classification result; S4, based on the training set, jointly optimize the model parameters of the location extraction task and the disaster classification task through a multi-task loss function; S5. After the multi-task loss function is jointly optimized and trained, input typhoon-related social media text data to extract and classify the location of typhoon disasters.

[0007] Preferably, step S1 specifically includes: Social media texts were collected using web crawlers with typhoon names and disaster keywords; the texts were segmented and converted into encoded sequences based on a vocabulary; special classification tags were added to the beginning and end of the encoded sequences, and the sequence length was standardized by zero padding; zero vectors were added to the end of shorter sentences to ensure consistent text length; and embedded information indicating the positional order of words in the text was added to obtain text vectors.

[0008] Preferably, the location extraction task in step S3 specifically includes: The model uses a BiLSTM layer to capture text sequence features bidirectionally and outputs an emission score matrix for each character. Based on the emission fraction matrix, the CRF layer of the model is used to calculate the BIO tag sequence through the transition fraction matrix and the Viterbi algorithm, and the place name entities in the BIO tag sequence are concatenated into complete place name entities; The complete place name entities are converted into latitude and longitude coordinates through the API interface of the map application, and then parsed into provincial and municipal place name administrative information through reverse geocoding.

[0009] Preferably, the place name enhancement weighting mechanism in step S3 includes: Map the obtained BIO tag sequence to a weight vector:

[0010] in, H Represents the mapped vector. For trainable weight matrix, For bias terms; The place name enhancement weight for each character is calculated using the Softmax function:

[0011] in, Representing the Place names with individual characters have increased weight. For learnable parameters, This is the hidden representation of the character 'i'. The weight vector is multiplied by the text representation to enhance the feature contribution of place-name related characters in disaster classification, and then fed into the CNN model:

[0012] in, This is a text representation with enhanced place name weights. = This adds weight vectors to place names, where ⊙ represents element-wise multiplication.

[0013] Preferably, the multi-task loss function in step S5 Loss extracted from location and disaster identification and loss The composition includes the loss term of the auxiliary classifier, and the calculation formula is as follows:

[0014]

[0015]

[0016] in, 1 and The cross-entropy loss is calculated using the two auxiliary classifiers. FocalLoss for CNN models; and These represent the weight factors for the auxiliary classifier and multi-task learning, respectively, and are set to 0.3 and 0.001.

[0017] Preferably, the location extraction loss The probability of predicting the label sequence is calculated, and the negative log-likelihood function is used to measure the degree of matching between the predicted sequence and the true sequence, i.e., the loss function. The calculation is as follows:

[0018]

[0019] in, It is a text representation. It is the predicted label sequence. It is a real label sequence. It represents all possible label sequences.

[0020] Preferably, the disaster classification task adopts a multi-label classification mechanism, and the output includes at least the disaster categories of traffic impact, social impact, power impact, forestry impact and waterlogging. If a single text involves multiple disaster categories, all relevant categories are labeled simultaneously. The CNN model uses multi-size convolutional kernels to extract features. After max pooling, the pooling results are concatenated and mapped to the label space through a fully connected layer. The probability of each label category in the label space is output through the Sigmoid function.

[0021] Preferably, the method further includes a disaster spatiotemporal visualization step S6: The extracted latitude and longitude coordinates are associated with disaster categories and a disaster spatial distribution map is generated based on a geographic information system, the number of disasters is counted according to a time sequence, and a time evolution map is generated.

[0022] To achieve the object of the present application, the second aspect provides a typhoon disaster location extraction and classification system based on social media, which applies the typhoon disaster location extraction and classification method based on social media described above, and the system comprises: A data acquisition and preprocessing module is configured to acquire social media text data and generate a text vector; A text representation module comprises a BERT model and an auxiliary classifier, and is configured to output a semantic vector fused with multiple hidden layers; A location extraction module comprises a BiLSTM-CRF model, and is configured to output a place name coordinate and place name administrative information; A disaster identification module comprises a CNN model, and is configured to output a multi-label disaster classification result; A multi-task learning module is configured to dynamically fuse location and disaster semantic features through a place name enhanced weight mechanism and jointly optimize model parameters of other modules through a multi-task loss function; Preferably, the system further comprises a spatio-temporal visualization module configured to generate a disaster spatial distribution map and a time evolution map.

[0023] Compared with the prior art, the present application has the following beneficial effects: The present application enhances the understanding ability of the model based on a multi-task learning framework, and synchronously optimizes the model of the location extraction and disaster classification tasks through a multi-task loss function, thereby improving the comprehensiveness and accuracy of the extraction and classification of the location information and disaster categories in the typhoon disaster assessment. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A step flowchart of the typhoon disaster location extraction and classification method based on social media provided by the present application is provided; Figure 2 A structure schematic diagram of the text representation module provided by the present application is provided; Figure 3 A structure schematic diagram of the location extraction module provided by the present application is provided; Figure 4 A structure schematic diagram of the disaster identification module provided by the present application is provided; Figure 5 A disaster time mode evolution map of three typhoons (i.e., “Yanhuo”, “Sianbal”, and “Dusery”) landing in the coastal areas of China is provided. DETAILED DESCRIPTION

[0025] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present application, but not to limit the scope of the present application.

[0026] Example 1 Please refer to Figure 1 The embodiment provides a typhoon disaster location extraction and classification method based on social media, comprising the following steps: S1, collecting typhoon-related social media text data, and preprocessing the text to generate a text vector and divide it into a training set; S2, extracting the multi-hidden layer semantic representation of the text vector through the BERT model, and performing fusion to generate a comprehensive text representation vector, using an auxiliary classifier to receive the output of each hidden layer to generate an intermediate result of location extraction and disaster classification to speed up subsequent model training; S3, synchronously performing a location extraction task and a disaster classification task based on a multi-task learning framework; The location extraction task identifies place entities and outputs a BIO label sequence through a BiLSTM-CRF model, completing location extraction; The disaster classification task dynamically calculates the weight of the place entity in the text through a place-enhanced weight mechanism, strengthens the semantic features related to the location in the disaster classification task, performs multi-label classification through a CNN model, and outputs the disaster classification result; S4, based on the training set, jointly optimizing the model parameters of the location extraction task and the disaster classification task through a multi-task loss function; S5, after the joint optimization training of the multi-task loss function is completed, inputting the typhoon-related social media text data to perform typhoon disaster location extraction and classification.

[0027] The typhoon disaster location extraction and classification method based on social media in Example 1 further comprises a disaster spatio-temporal visualization step S6: associating the extracted latitude and longitude coordinates with the disaster category and generating a disaster spatial distribution map based on a geographic information system, and generating a disaster time pattern evolution map by statistically analyzing the number of disasters in a time sequence.

[0028] Step S1 in Example 1 specifically comprises: collecting social media text through network crawlers with typhoon names and disaster keywords; performing word segmentation processing on the text, converting characters into encoded sequences according to a vocabulary; adding special classification markers at the beginning and end of the encoded sequence, and uniformly filling the sequence length by zero padding; filling zero vectors at the end of the text with shorter sentence length to make the length of each text consistent; increasing embedded information representing the order of words in the text to obtain a text vector.

[0029] The multi-task learning framework of the social media-based typhoon disaster location extraction and classification method of embodiment 1 is specifically implemented by four parts: text representation (TR), location extraction (LE), disaster identification (DI), and multi-task learning (MTL), as shown in FIGS. 3, Figure 2 、 3 , 4. First, the text data from the social media platform is converted into a text vector by the TR module. Further, the LE module and the DI module extract the location information and the disaster category-related semantic features, respectively. Further, the MTL module is used for feature fusion to enhance the text representation. Finally, the multi-task loss function is used to realize the training optimization of the text location information and the disaster category extraction and classification.

[0030] The TR module is composed of BERT and an auxiliary classifier. BERT adopts an architecture composed of 12 layers of Transformer encoders, each of which can understand the context relationship in the text. A key feature is the bidirectional attention mechanism, which allows the model to consider the context information of each word in the sentence simultaneously. Multi-head attention further enhances the bidirectional understanding of the model, which helps to capture different meanings in the text. In addition, since the outputs of different hidden layers of BERT capture various semantic information, the present application not only uses the output of the last hidden layer, but also supplements the information by integrating the outputs of other layers. These hidden layer text representations are input into the auxiliary classifier to generate the outputs of location extraction and disaster category classification, thereby improving the parameter update during model training.

[0031]

[0032] wherein, represents the outputs of location extraction and disaster classification, represents the auxiliary classifier, is the hidden layer text representation.

[0033] The LE module is composed of BiLSTM (Bidirectional Long Short-Term Memory) and CRF (Conditional Random Field) and builds a place name entity recognition model to extract location information from the text. BiLSTM is divided into two layers: the forward LSTM processes the input from left to right, while the backward LSTM processes in the opposite direction. Each layer includes cell units responsible for maintaining the state of the unit and three gates (i.e., forget gate, input gate, and output gate), which regulate the information processing at each time step. These gates decide whether to remember, forget, or output the information in the text during the processing of the model. The forget gate ( ), the input gate ( ), and the output gate ( ) calculation formulas are as follows:

[0034]

[0035]

[0036] where, , , is a weight matrix; , , is a bias vector; is the hidden state of the previous time step; is the input word vector of the current time step; is a sigmoid function that controls the value of the gate between 0 and 1.

[0037] The temporary cell state (h ) represents the new information input to the cell state, and the calculation formula is:

[0038] where, is a weight matrix; is a bias vector; is a hyperbolic tangent function.

[0039] Next, update the cell state (h ) and hidden state (h ) of the current time step, the calculation formula is as follows:

[0040]

[0041] where, is the cell state of the previous time step BiLSTM enhances semantic understanding of text representation by capturing information from two directions in the text. The output of BiLSTM is passed through a fully connected layer to generate an emission score matrix of size n x k, where n is the number of characters and k is the number of place name labels. This matrix reflects the likelihood of each word for each label. Then the sequence score is calculated as the sum of all word emission scores and transition scores.

[0042]

[0043] where, is the text representation, is the predicted label sequence; represents the emission score matrix, represents the word corresponding label score of the label denotes the transition score matrix, denotes the score from label to label .

[0044] CRF utilizes these emission and transition scores to model the relationship between adjacent labels, and adopts the Viterbi algorithm to determine the most likely label sequence, which can be calculated as follows:

[0045] where, is the true label sequence, is all possible label sequences, is the highest score output sequence.

[0046] After assigning BIO labels to each character in the text, the location is geocoded to the corresponding coordinates through the Gaode Map application interface, and then the associated province and city are determined using reverse geocoding.

[0047] The DI module is composed of CNN and builds a multi-label classification model to realize the classification of disaster categories from the text. CNN adopts multiple convolution kernels of different sizes to capture different features in the text, among which smaller kernels focus on character-level patterns, and larger kernels capture higher-level semantic patterns. After the convolution operation, a max-pooling layer is applied to downsample the feature maps generated by each kernel, reducing the dimension while preserving the basic information. Then these pooled features are concatenated and flattened into a single vector, and the probability of the text corresponding to different category labels is calculated using the sigmoid function to determine its disaster category.

[0048] The MTL module is composed of place name enhancement weights and multi-task loss functions. Among them, the place name enhancement weights further strengthen the text representation by establishing a connection between the location extraction and disaster category classification tasks. First, the BIO label sequence in the location extraction task B is mapped to the place name weight space through the fully connected layer to learn the relationship between place name entities and non-place name entities, which can be represented as:

[0049] where, H denotes the mapped vector, is a trainable weight matrix, is a bias term.

[0050] Then, the Softmax function is used to calculate the place name enhancement weight of each word, which represents its importance in the disaster classification task, and the calculation formula is as follows:

[0051] where, represents the place name enhanced weight of the i-th character, is a learnable parameter, is the hidden layer representation of character i. In the disaster identification task, the place name enhanced weight is multiplied with the text representation, so that the words related to the place name are strengthened in the feature extraction process, and the sensitivity of the classification model to the disaster related location information is improved, which can be expressed as:

[0052]

[0053] where, is the text representation after the place name weight enhancement, = is the place name enhanced weight vector, and represents element-wise multiplication.

[0054] In the subsequent convolutional layer and max pooling layer, the weight information is constantly updated to highlight the key location information in the disaster text, thereby effectively balancing the location extraction and disaster classification tasks and improving the overall performance of the model.

[0055] The multi-task loss function is composed of the losses of the LE module and the DI module. For the place name entity recognition model in the LE module, the probability of the predicted label sequence is calculated, and the negative log likelihood function is used to measure the matching degree of the predicted sequence and the true sequence, and the loss function is calculated as follows:

[0056]

[0057] For the multi-label classification model in the DI module, Focal Loss is used as the loss function to enhance the learning ability of unbalanced classes, which is represented as In addition, the outputs of the 4th and 8th layers of BERT are introduced as two auxiliary classifiers to further enhance the text representation and speed up the model training. The total loss of the model is composed of the location extraction loss and the disaster identification loss , and also includes the loss term of the auxiliary classifier, whose calculation formula is as follows:

[0058]

[0059]

[0060] ​wherein, 1 and are the losses of two auxiliary classifiers. and respectively represent the weight factors of auxiliary classifiers and multi-task learning, which are set to 0.3 and 0.001 respectively in the experiment.

[0061] Embodiment 2 The embodiment 2 provides a typhoon disaster location extraction and classification system based on social media, applying the typhoon disaster location extraction and classification method based on social media in embodiment 1, and the system comprises: a data acquisition and preprocessing module, which is used for acquiring social media text data and generating a text vector; a text representation module, which comprises a BERT model and an auxiliary classifier, and is used for outputting a semantic vector fused with multiple hidden layers; a location extraction module, which comprises a BiLSTM-CRF model, and is used for outputting a place name coordinate and place name administrative information; a disaster identification module, which comprises a CNN model, and is used for outputting a multi-label disaster classification result; a multi-task learning module, which is used for dynamically fusing location and disaster semantic features through a place name enhanced weight mechanism and jointly optimizing and training model parameters of other modules through a multi-task loss function.

[0062] The system in the embodiment 2 further comprises a spatio-temporal visualization module, which is used for generating a disaster spatial distribution map and a disaster time pattern evolution atlas.

[0063] Embodiment 3 The embodiment 3 is based on the embodiments 1 and 2, and further illustrates the specific implementation of the present application.

[0064] The present case is based on three typhoons (i.e. “Yan Hua”, “Siam Ba” and “Du Su Rui”) landing in the coastal areas of China, and the social media text data is taken as an example to evaluate the typhoon disaster, as shown in Figure 5 .

[0065] Step 1: Data preparation. Through the network crawler technology to simulate the browser behavior for automatic data collection, the typhoon name and the disaster-related terms are set as search keywords, and the microblog text within 7 days before and after the typhoon landing is collected. The obtained data fields include username, text content, posting time, user registration location and geographic marker location.

[0066] Second step: The microblog text needs to be pre-processed and converted into a text vector before classification. Microblog text preprocessing includes: word segmentation and coding: the text is divided into individual words by a word segmenter, and the words are converted into corresponding codes according to the vocabulary; adding markers: adding special classification markers at the beginning and end of the coded text; filling in the blanks: filling in the zero vector at the end of the text if the sentence length is short, so that the length of each text is consistent; position embedding: adding embedding information representing the position sequence of words in the text. The pre-processed text vector of the microblog text will be used as the input of the model.

[0067] Third step: According to the disaster report standard of "Meteorological Disaster Investigation Technical Specification Meteorological Disaster Information Collection", and combined with the ability of the text to reflect the disaster, the typhoon disaster is divided into five categories: traffic impact, social impact, power impact, forestry impact and waterlogging. Based on the typhoon disaster classification system, the positive samples are labeled with multiple labels, that is, if a text contains information of multiple disaster categories, the text is labeled as multiple categories at the same time.

[0068] Fourth step: Construct the SMBTDLCMS model, and evaluate and analyze the spatial and temporal characteristics of different typhoon disasters based on the SMBTDLCMS model. In the experiment, 70% of the data is randomly selected as the training set, 15% as the validation set, and 15% as the test set. After 30 rounds of training, the pth file is saved. The parameters are used as model parameters to predict the results of all texts. Based on Python and ArcGIS, the final results are obtained.

[0069] In summary, according to the above embodiment, the engineering design of the social media-based typhoon disaster location and classification method and system SMBTDLCMS successfully improves the usability of the model in practical typhoon disaster assessment while maintaining the model accuracy. The model is based on a multi-task learning framework to enhance the understanding ability of the model, making this engineering design perform better than other models in both location extraction and disaster category classification tasks. The experimental results of the invention show high performance, with an F1 score of 0.891 for location extraction and an F1 score of 0.898 for disaster identification. Through case studies of three typhoons that landed in the coastal areas of China (namely "Yan Hua", "Siam Ba", and "Du Su Rui"), the spatial distribution and temporal pattern of typhoon disasters are revealed, providing actionable insights for disaster management and resource allocation. This framework not only advances typhoon disaster assessment, but also provides broad applicability for other disaster scenarios, helping to enhance the resilience and sustainable development of cities.

[0070] The above describes the preferred embodiments of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, several improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A method for extracting and classifying typhoon disaster locations based on social media, characterized in that, Includes the following steps: S1. Collect typhoon-related social media text data, preprocess the text to generate text vectors, and divide them into training sets; S2. Extract multi-hidden-layer semantic representations of text vectors through the BERT model, fuse them to generate comprehensive text representation vectors, use an auxiliary classifier to receive the outputs of each hidden layer, and generate intermediate results for location extraction and disaster classification to accelerate subsequent model training. S3. Based on a multi-task learning framework, the location extraction task and the disaster classification task are executed simultaneously. The location extraction task identifies place name entities and outputs BIO label sequences through a BiLSTM-CRF model to complete the location extraction. The disaster classification task dynamically calculates the weights of place name entities in the text through a place name enhancement weight mechanism to strengthen the semantic features related to location in the disaster classification task. Multi-label classification is performed through a CNN model to output the disaster classification results. S4. Based on the training set, jointly optimize the model parameters for the training location extraction task and the disaster classification task through a multi-task loss function; S5. After the multi-task loss function is jointly optimized and trained, input typhoon-related social media text data to extract and classify the location of typhoon disasters.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: Social media texts were collected using web crawling technology based on typhoon names and disaster keywords; the texts were segmented and converted into encoded sequences according to a vocabulary; special classification tags were added to the beginning and end of the encoded sequences, and the sequence length was standardized by zero padding; zero vectors were added to the end of shorter texts to ensure that the length of each text was consistent; and embedded information indicating the positional order of words in the text was added to obtain text vectors.

3. The method according to claim 1, characterized in that, The location extraction task in step S3 specifically includes: The model uses a BiLSTM layer to capture text sequence features bidirectionally and outputs an emission score matrix for each character. Based on the emission fraction matrix, the CRF layer of the model is used to calculate the BIO tag sequence through the transition fraction matrix and the Viterbi algorithm, and the place name entities in the BIO tag sequence are concatenated into complete place name entities; The complete place name entities are converted into latitude and longitude coordinates through the API interface of the map application, and then parsed into provincial and municipal place name administrative information through reverse geocoding.

4. The method according to claim 3, characterized in that, The place name enhancement weighting mechanism in step S3 includes: Map the obtained BIO tag sequence to a weight vector: in, H Represents the mapped vector. For trainable weight matrix, For bias terms; The place name enhancement weight for each character is calculated using the Softmax function: in, Representing the Place names with individual characters have increased weight. For learnable parameters, This is the hidden representation of the character 'i'. The weight vector is multiplied by the text representation to enhance the feature contribution of place-name related characters in disaster classification, and then fed into the CNN model: in, This is a text representation with enhanced place name weights. = This adds weight vectors to place names, where ⊙ represents element-wise multiplication.

5. The method according to claim 1, characterized in that, The multi-task loss function in step S5 Loss extracted from location and disaster identification and loss The composition includes the loss term of the auxiliary classifier, and the calculation formula is as follows: in, 1 and The cross-entropy loss is calculated using the two auxiliary classifiers. FocalLoss for CNN models; and These represent the weight factors for the auxiliary classifier and multi-task learning, respectively, and are set to 0.3 and 0.

001.

6. The method according to claim 5, characterized in that, Location extraction loss The probability of predicting the label sequence is calculated, and the negative log-likelihood function is used to measure the degree of matching between the predicted sequence and the true sequence, i.e., the loss function. The calculation is as follows: in, It is a text representation. It is the predicted label sequence. It is a real label sequence. It represents all possible label sequences.

7. The method according to claim 1, characterized in that, The disaster classification task adopts a multi-label classification mechanism, and the output includes at least the disaster categories of traffic impact, social impact, power impact, forestry impact and waterlogging. If a single text involves multiple disaster categories, all relevant categories are labeled simultaneously. The CNN model uses multi-size convolutional kernels to extract features. After max pooling, the pooling results are concatenated and mapped to the label space through a fully connected layer. The probability of each label category in the label space is output through the Sigmoid function.

8. The method according to claim 1, characterized in that, The method also includes a disaster spatiotemporal visualization step S6: The extracted latitude and longitude coordinates are associated with disaster categories, and a spatial distribution map of disasters is generated based on a geographic information system. The changes in the number of disasters are statistically analyzed according to time series, and a disaster time pattern evolution map is generated.

9. A typhoon disaster location extraction and classification system based on social media, employing the typhoon disaster location extraction and classification method based on social media as described in any one of claims 1-8, characterized in that, include: Data acquisition and preprocessing module: used to acquire social media text data and generate text vectors; Text representation module: Integrates BERT model and auxiliary classifier to output semantic vectors fused from multiple hidden layers; Multi-task learning module: includes a location extraction submodule based on BiLSTM-CRF model for outputting place name coordinates and place name administrative information, and a disaster recognition submodule based on CNN model for outputting multi-label disaster classification results; Feature enhancement module: used to dynamically fuse location and disaster semantic features through a place name enhancement weighting mechanism; Loss optimization module: Used to jointly optimize the model parameters for the training location extraction task and the disaster identification task through a multi-task loss function.

10. The system according to claim 9, characterized in that, The system also includes a spatiotemporal visualization module, used to generate spatial distribution maps of disasters and temporal pattern evolution maps of disasters.

Images