Disaster information pushing method and device based on large language model, equipment and readable medium

Abstract

Embodiments of the present disclosure disclose a disaster information pushing method and device based on a large language model, an equipment and a medium. A specific implementation of the method comprises: obtaining environmental text information; performing feature extraction processing based on the environmental text information to generate a plurality of extracted features and obtain an extracted feature group; classifying the extracted features; preprocessing at least one local feature and at least one global feature; inputting the local feature group and the global feature group into a pre-trained large language model to obtain disaster information and disaster response information corresponding to the disaster information; generating a disaster pushing page based on the disaster information and the disaster response information, and pushing the disaster pushing page to an associated user terminal for display. The implementation accurately determines the disaster level and avoids waste of personnel casualties or emergency response resources.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and specifically to a method, apparatus, device, and readable medium for disaster information push based on a large language model. Background Technology

[0002] The importance of meteorological emergency response in disaster prevention and mitigation is becoming increasingly prominent. Currently, the common approach to meteorological emergency response is to classify meteorological data using a trained convolutional neural network, determine the corresponding meteorological disaster level, and generate a corresponding emergency response.

[0003] However, when using the above methods for emergency response to meteorological conditions, the following technical problems often arise:

[0004] When using convolutional neural networks to classify meteorological and environmental data to determine disaster levels, the convolutional neural networks cannot effectively integrate multidimensional features to determine disaster levels for structured meteorological and environmental data, unstructured data, and geographic information, resulting in inaccurate disaster levels and causing casualties or waste of emergency response resources.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure propose a disaster information push method, apparatus, electronic device, and computer-readable medium based on a large language model to solve one or more of the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a disaster information push method based on a large language model. The method includes: acquiring environmental text information, wherein the environmental text information is text information used to characterize environmental information received from a target terminal through a preset interface; performing feature extraction processing based on the environmental text information to generate multiple extracted features, obtaining an extracted feature group; classifying each extracted feature in the extracted feature group to generate at least one local feature and at least one global feature; preprocessing the at least one local feature and the at least one global feature to generate preprocessed local features and global features, obtaining a local feature group and a global feature group; inputting the local feature group and the global feature group into a pre-trained large language model to obtain disaster information and corresponding disaster response information; generating a disaster push page based on the disaster information and the disaster response information; and pushing the disaster push page to an associated user terminal for display.

[0009] Secondly, some embodiments of this disclosure provide a disaster information push device based on a large language model. The device includes: an acquisition unit configured to acquire environmental text information, wherein the environmental text information is text information used to characterize environmental information received from a target terminal through a preset interface; an execution unit configured to perform feature extraction processing based on the environmental text information to generate multiple extracted features and obtain an extracted feature group; a classification unit configured to classify each extracted feature in the extracted feature group to generate at least one local feature and at least one global feature; a preprocessing unit configured to preprocess the at least one local feature and the at least one global feature to generate preprocessed local features and global features, obtaining a local feature group and a global feature group; an input unit configured to input the local feature group and the global feature group into a pre-trained large language model to obtain disaster information and disaster response information corresponding to the disaster information; and a generation unit configured to generate a disaster push page based on the disaster information and the disaster response information, and push the disaster push page to an associated user terminal for display.

[0010] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0011] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0012] The above embodiments of this disclosure have the following beneficial effects: The disaster information push method based on a large language model, as described in some embodiments of this disclosure, accurately determines the disaster level, avoiding casualties or waste of emergency response resources. Specifically, the reason for inaccurate disaster level determination and the resulting casualties or waste of emergency response resources is that when using convolutional neural networks to classify meteorological environmental data to determine the disaster level, the convolutional neural network cannot effectively fuse multi-dimensional features to determine the disaster level for structured data, unstructured data, and geographic information, leading to inaccurate disaster level determination and resulting in casualties or waste of emergency response resources. Based on this, the disaster information push method based on a large language model, as described in some embodiments of this disclosure, firstly, acquires environmental text information. This allows obtaining meteorological text data from the meteorological monitoring center. Secondly, based on the aforementioned environmental text information, feature extraction processing is performed to generate multiple extracted features, resulting in an extracted feature group. This allows the extraction of data features from the text data. Then, for each extracted feature in the extracted feature group, the extracted feature is classified to generate at least one local feature and at least one global feature. This allows the distinction between global and local features. Next, the at least one local feature and the at least one global feature are preprocessed to generate preprocessed local and global features, resulting in local feature sets and global feature sets. This allows for the preprocessing of local and global features. Then, the local and global feature sets are input into a pre-trained large language model to obtain disaster information and corresponding disaster response information. Thus, by fusing and analyzing structured data, unstructured data, and geographic information from the large language model's environment, the disaster level corresponding to the meteorological environment can be accurately determined, thereby avoiding casualties or waste of emergency response resources. Attached Figure Description

[0013] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0014] Figure 1 This is a flowchart of some embodiments of the model parameter tuning method based on a large language model according to the present disclosure;

[0015] Figure 2 These are schematic diagrams illustrating the structure of some embodiments of the model parameter tuning device based on a large language model according to this disclosure;

[0016] Figure 3 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0017] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0018] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0019] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0020] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0021] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0022] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] Figure 1 A flowchart 100 is shown, illustrating some embodiments of a disaster information push method based on a large language model according to this disclosure. This disaster information push method based on a large language model includes the following steps:

[0024] Step 101: Obtain environmental text information.

[0025] In some embodiments, the execution entity (e.g., a server) of the disaster information push method based on a large language model can acquire environmental text information. This environmental text information is text information representing environmental information received from a target terminal through a preset interface. The preset interface can be a pre-defined interface for acquiring environmental text information from the target terminal. The target terminal is a meteorological terminal connected to a meteorological bureau.

[0026] Step 102: Based on the environmental text information, perform feature extraction processing to generate multiple extracted features and obtain an extracted feature group.

[0027] In some embodiments, the execution entity may perform feature extraction processing based on the aforementioned environmental text information to generate multiple extracted features, thereby obtaining an extracted feature group. The feature extraction processing may involve extracting various features from the aforementioned environmental text information using an attention mechanism.

[0028] In the process of adopting technical solutions to address the technical problems mentioned in the background, the following technical issues often arise: some meteorological monitoring equipment is deployed in mountainous areas with poor communication conditions, resulting in high delays in meteorological data transmission, making it impossible to determine the disaster level in a timely manner, leading to casualties or wasted emergency response resources. Conventional solutions to these technical problems generally involve enhancing and improving the data transmission equipment corresponding to the meteorological monitoring equipment in mountainous areas, or erecting more signal towers to reduce data transmission delays. However, these conventional solutions still have the following problems: enhancing and improving data transmission equipment, or erecting more signal towers, results in high data transmission costs. Considering the problems of the conventional solutions and the aforementioned technical issues, and in light of the current technological situation, the following solution can be adopted.

[0029] Optionally, after step 102, the following steps may also be included:

[0030] The first step is to determine the network status corresponding to the above environmental text information.

[0031] In some embodiments, the executing entity can determine the network status corresponding to the environmental text information. Here, the network status of the meteorological monitoring equipment corresponding to the environmental text information can be determined as the network status corresponding to the environmental text information.

[0032] The second step is to load a preset feature configuration information set in response to the network state meeting the first preset condition.

[0033] In some embodiments, the execution entity may load a preset feature configuration information set in response to the network state meeting a first preset condition. The preset feature configuration information set may be a pre-set weight value for each extracted feature. The first preset condition may be a network interruption state.

[0034] The third step is to generate a meteorological emergency assessment corresponding to the aforementioned extracted feature set based on the aforementioned preset feature configuration information set and the aforementioned extracted feature set.

[0035] In some embodiments, the executing entity can generate a meteorological emergency assessment corresponding to the aforementioned extracted feature set based on the aforementioned preset feature configuration information set and the aforementioned extracted feature set. In practice, the product of each extracted feature in the aforementioned extracted feature set and its corresponding preset feature configuration information can be determined as a sub-meteorological emergency assessment, and the sum of the normalized sub-meteorological emergency assessments can be determined as the meteorological emergency assessment.

[0036] The fourth step involves generating a meteorological emergency warning based on the aforementioned meteorological emergency assessment, and then pushing the meteorological emergency warning to the relevant user terminals.

[0037] In some embodiments, the aforementioned implementing entity may generate a meteorological emergency warning based on the aforementioned meteorological emergency assessment, and push the aforementioned meteorological emergency warning to the associated user terminal.

[0038] The fifth step involves selecting each extracted feature that meets the preset query conditions from the above-mentioned extracted feature set based on the historical query information set, and marking each selected extracted feature to generate marked extracted features.

[0039] In some embodiments, the execution entity may select each extracted feature that meets a preset query condition from the extracted feature set based on the historical query information set, and mark each selected extracted feature to generate marked extracted features. The preset query condition may be that the number of historical query information corresponding to the extracted feature is greater than or equal to a preset query quantity.

[0040] Step 6: In response to the detection that the network status meets the second preset condition, determine the network transmission speed corresponding to the above network status.

[0041] In some embodiments, the execution entity may determine the network transmission speed corresponding to the network state in response to detecting that the network state meets a second preset condition. The second preset condition may be a network connectivity status.

[0042] Step 7: Based on the network transmission speed mentioned above, select a preset compression algorithm corresponding to the network transmission speed from the preset compression algorithm library, and compress the feature extraction set of each tag and the feature extraction set after removing the feature extraction of each tag.

[0043] In some embodiments, the execution entity may select a preset compression algorithm corresponding to the network transmission speed from a preset compression algorithm library, based on the network transmission speed, and compress the extracted features of each marker and the extracted feature set after removing the extracted features of each marker. Here, the preset compression algorithm library may be a pre-defined algorithm database for storing compression algorithms.

[0044] Step 8: Transmit the compressed marker features to the target server according to the preset transmission protocol.

[0045] In some embodiments, the executing entity may transmit the compressed tag-extracted features to the target server according to a preset transmission protocol. The preset transmission protocol may be an incremental update protocol.

[0046] Steps one through eight above, as an inventive point of this disclosure, combined with optional steps below, solve the technical problem: "Some meteorological monitoring equipment is deployed in mountainous areas with poor communication conditions, resulting in high delays in meteorological data transmission, making it impossible to determine the disaster level in a timely manner, causing casualties or wasting emergency response resources." The reasons for this casualty or wasted emergency response resources are as follows: Some meteorological monitoring equipment is deployed in mountainous areas with poor communication conditions, resulting in high delays in meteorological data transmission, making it impossible to determine the disaster level in a timely manner, causing casualties or wasting emergency response resources. Solving these factors can avoid casualties or wasted emergency response resources. To achieve this effect, this disclosure firstly determines the network status corresponding to the aforementioned environmental text information. This allows determination of whether the current network can transmit data in real time. Secondly, in response to the aforementioned network status meeting a first preset condition, a preset feature configuration information set is loaded; based on the aforementioned preset feature configuration information set and the aforementioned extracted feature set, a meteorological emergency evaluation corresponding to the aforementioned extracted feature set is generated. Thirdly, based on the aforementioned meteorological emergency evaluation, a meteorological emergency warning is generated, and the aforementioned meteorological emergency warning is pushed to associated user terminals. Therefore, temporary emergency warnings can be generated and pushed out based on edge computing. Fourth, based on the historical query information set, each extracted feature that meets the preset query conditions is selected from the above extracted feature set, and each selected extracted feature is marked to generate marked extracted features. This allows for the selection of extracted features with high query frequency. Fifth, in response to the detection that the network state meets the second preset condition, the network transmission speed corresponding to the above network state is determined; based on the above network transmission speed, a preset compression algorithm corresponding to the above network transmission speed is selected from the preset compression algorithm library, and each marked extracted feature and the extracted feature set after removing each marked extracted feature are compressed; the compressed marked extracted features are transmitted to the target server according to the preset transmission protocol. This allows for the selection of different compression algorithms based on different network transmission speeds, thereby enabling faster transmission of extracted features to the server for more accurate calculations. Combined with the optional steps below, based on the above disaster response information, the associated emergency response equipment is controlled to execute the corresponding disaster emergency response tasks. This reduces the negative impact of network latency, thereby avoiding casualties or waste of emergency response resources.

[0047] Step 103: For each extracted feature in the extracted feature group, the extracted feature is classified to generate at least one local feature and at least one global feature.

[0048] In some embodiments, the execution entity may classify each extracted feature in the extracted feature group to generate at least one local feature and at least one global feature.

[0049] Step 104: Preprocess at least one local feature and at least one global feature to generate preprocessed local features and global features, resulting in a local feature group and a global feature group.

[0050] In some embodiments, the execution entity may preprocess the at least one local feature and the at least one global feature to generate preprocessed local features and global features, thereby obtaining a local feature group and a global feature group.

[0051] In practice, preprocessing can be performed through the following steps:

[0052] The first step is to perform synonym replacement on the minority classes of at least one local feature and at least one global feature.

[0053] The second step is to add noise to each local feature and each global feature to generate local features and global features with added noise.

[0054] Step 105: Input the local feature group and the global feature group into the pre-trained large language model to obtain disaster information and disaster response information corresponding to the above disaster information.

[0055] In some embodiments, the executing entity can input the aforementioned local feature sets and global feature sets into a pre-trained large language model to obtain disaster information and corresponding disaster response information. The aforementioned large language model includes an input processing layer, an encoder layer, a decoder layer, and an output layer.

[0056] In practice, the above output layer can be trained through the following steps:

[0057] The first step involves receiving at least one output vector from the decoder within the large language model. This decoder can be composed of 28 Transformer layers, each including a self-attention mechanism and a feedforward neural network. Each output vector contains dimensionality information, which characterizes the feature dimensions.

[0058] The second step involves enhancing at least one output vector based on the aforementioned feature fusion module to generate an enhanced output vector set. This enhancement process can be a feature fusion process. The feature fusion module includes local feature units and global feature units.

[0059] In practice, at least one of the above output vectors can be enhanced using the following steps:

[0060] The first step involves inputting at least one of the output vectors into the local feature units included in the feature fusion module to obtain the local feature set corresponding to the output vector. In practice, local features can be obtained using the following formula:

[0061] a i =softmax(W q ·h i ).

[0062] Among them, a i Let W represent the i-th local feature. q This represents a learnable 4096×4096-dimensional query matrix. i This represents the i-th output vector.

[0063] The second step is to generate a global feature set based on the aforementioned local feature sets and global feature units. This global feature set can be obtained through a weighted approach. In practice, the global features can be generated using the following formula:

[0064] g=∑a i ·h i .

[0065] Where g represents the global feature.

[0066] The third step involves selecting the corresponding global feature from the global feature set for each local feature in the aforementioned local feature set, and then fusing the local and global features to generate a fused feature, which serves as the enhanced output vector. The global and local features can be fused using the following formula:

[0067] f = LayerNorm(W f ·[h;g]+b f ).

[0068] Among them, W f This represents a fusion matrix of 4096×8192 dimensions. (b) f This represents a 4096-dimensional bias vector. f represents the enhanced output vector. [h; g] can be concatenated features obtained by splicing local and global features.

[0069] In the process of adopting technical solutions to solve the above-mentioned technical problems, the following technical problems often arise: when splicing and fusing local features with flattened or pooled global features at the channel dimension, redundancy may be introduced, resulting in the fused features not matching the actual required features, which in turn leads to inaccurate determination of disaster level, causing casualties or waste of emergency response resources.

[0070] Optionally, the above local features and the above global features can be fused using the following sub-steps:

[0071] The first sub-step involves upsampling the aforementioned global features to generate sampled global features. Here, bilinear interpolation or transposed convolution can be used for upsampling.

[0072] The second sub-step is to determine the number of feature channels for the aforementioned global features and local features.

[0073] The third sub-step, in response to the determination that the number of feature channels of the sampled global features and the sampled local features are inconsistent, adjusts the number of channels corresponding to the sampled global features to make the number of channels of the sampled global features and the sampled local features consistent. Here, the number of channels of the sampled global features can be adjusted by a 1×1 convolution.

[0074] The fourth sub-step involves performing a first convolution on the aforementioned local features to generate the local features after the first convolution. In practice, a 1×1 convolution can be performed on the aforementioned local features.

[0075] The fifth sub-step involves adding the local features after the first convolution and the global features after sampling element by element to generate the summed features.

[0076] The sixth sub-step involves performing a second convolution on the summed features to generate a fused convolutional feature, thus obtaining the convolutional feature. In practice, a 3×3 convolution can be used for the second convolution to obtain the convolutional feature.

[0077] The seventh sub-step involves batch normalizing the convolutional features to generate batch-normalized convolutional features, which serve as the enhanced output vector.

[0078] The first to seventh sub-steps described above, as an inventive point of this disclosure, combined with the optional steps below, solve the technical problem: "When fusing local features with flattened or pooled global features at the channel dimension, redundancy may be introduced, causing the fused features to be inconsistent with the actual required features, resulting in inaccurate determination of the disaster level and waste of personnel casualties or emergency response resources." The reasons for this waste of personnel casualties or emergency response resources are as follows: When fusing local features with flattened or pooled global features at the channel dimension, redundancy may be introduced, causing the fused features to be inconsistent with the actual required features, resulting in inaccurate determination of the disaster level and waste of personnel casualties or emergency response resources. If the above factors are resolved, the waste of personnel casualties or emergency response resources can be avoided. To achieve this effect, this disclosure firstly, upsamples the aforementioned global features to generate sampled global features. This increases the resolution of the global features. Second, determine the number of feature channels for the aforementioned global and local features. In response to the discrepancy between the number of feature channels for the sampled global and local features, adjust the number of channels for the sampled global features to make them consistent. This ensures that the number of channels for the global and local features is aligned. Third, perform a first convolution on the local features to generate convolutionally-enhanced local features. Add the convolutionally-enhanced local features and the sampled global features element-wise to generate a summed feature. This allows for the fusion of global and local features through element-wise addition, thereby enhancing the response of key regions. Fourth, perform a second convolution on the summed feature to generate a convolutionally-enhanced fused feature. Further convolution can then be applied to the fused feature to integrate local details and global semantics, generating the final feature map. Fifth, perform batch normalization on the convolutionally-enhanced feature to generate batch-normalized convolutionally-enhanced features, which serve as the enhanced output vector. This completes the fusion of local and global features, and batch normalization is used to improve the model's robustness. Combined with step 106 below, a disaster notification page is generated based on disaster information and disaster response information, and then pushed to relevant user terminals for display. This allows for accurate determination of the disaster level, thereby avoiding casualties or waste of emergency response resources.

[0079] The fourth step is to determine the generated enhanced output vectors as the enhanced output vector set.

[0080] The third step involves using the feature fusion module described above to perform noise suppression processing on each enhanced output vector in the enhanced output vector set to generate a suppressed enhanced output vector, which serves as the suppressed output feature to obtain the suppressed output feature set.

[0081] In some optional implementations of certain embodiments, the aforementioned execution entity may perform noise suppression processing through the following steps:

[0082] First, for each enhanced output vector in the above enhanced output vector set, perform the following suppression step:

[0083] The first suppression step involves determining the feature importance value corresponding to the enhanced output vector based on the gating unit included in the feature fusion module. Here, the feature importance value corresponding to the enhanced output vector can be determined using the following formula:

[0084] n=σ(W n ·f+b n ).

[0085] Among them, W n This represents a 4096×4096 dimensional transformation matrix. n represents the bias term. f represents the enhanced output vector.

[0086] The second suppression step, based on the aforementioned feature importance, performs feature filtering on the enhanced output vector to generate a filtered output vector. This feature filtering process can be expressed as f′=g⊙f.

[0087] The third suppression step involves combining the enhanced output vector and the filtered output vector to generate a combined vector, which serves as the suppressed output feature. This combined processing can be represented as o = f + f′.

[0088] Step 104: Based on the hierarchical classification module, classify each suppressed output feature in the suppressed output feature set to generate a classified output feature set.

[0089] In some embodiments, the execution entity may classify each suppressed output feature in the suppressed output feature set based on the hierarchical classification module to generate a classified output feature set. Each classified output feature in the classified output feature set corresponds to classification information. The hierarchical classification module includes a main classifier and a fine-grained classifier.

[0090] In practice, the following steps can be used to classify the various suppressed output features in the above-mentioned suppressed output feature set:

[0091] The first step is to perform the following processing steps for each suppressed output feature in the above suppressed output feature set:

[0092] The first processing step involves dimensionality reduction of the suppressed output features to generate dimensionality-reduced features. In practice, this dimensionality reduction can be achieved using the linear layers included in the main classifier of the hierarchical classification module.

[0093] FC1(x) = ReLU(W1·x+b1).

[0094] Where W1 represents a 2048×4096 dimensional matrix. x represents the suppressed output feature. b1 represents the bias term in the FC1 function.

[0095] The second processing step involves mapping the aforementioned dimensionality-reduced features to a preset category space to generate mapped features. This mapping can be achieved using the following formula:

[0096] FC2(FC1(x))=W2·FC1(x)+b2.

[0097] Here, W2 represents a custom dimension matrix, the dimensions of which are set according to the number of main categories. b2 represents the bias term in the FC2 function.

[0098] The third processing step involves performing fine-grained classification on the mapping features based on the fine-grained classifier included in the hierarchical classification module, to generate fine-grained classified mapping features as output features after classification.

[0099] The fourth step involves inputting the classified feature set into the loss function module to obtain the feature loss value set. The loss function module includes: a main classification loss function, a fine-grained classification loss function, and a feature consistency loss function.

[0100] In some optional implementations of certain embodiments, the main classification loss function, fine-grained classification loss function, and feature consistency loss function described above can be expressed by the following formula:

[0101] Main classification loss function:

[0102]

[0103] Here, α can be set to 0.25. γ can be initially set to 2. p represents the probability value predicted by the model. Here, the class weight α is smoothed by an exponential function using the reciprocal of the class frequency as the initial weight: Normalization ensures the weights sum to 1. γ increases with each training epoch. Used to adaptively adjust the weights of difficult samples.

[0104] Fine-grained classification loss function:

[0105]

[0106] Where ∈ can be set to 0.1. y true This represents the actual label. N represents the total number of categories.

[0107] Feature consistency loss function:

[0108]

[0109] Among them, F local and F global These represent the probability distributions of local features and global features, respectively.

[0110] Total loss through weighted combination:

[0111]

[0112] The weight coefficients in the total loss mentioned above were determined through grid search optimization.

[0113] The fifth step is to optimize and tune the parameters of the large language model based on the above feature loss value set.

[0114] Alternatively, the above output layer can be trained using the following strategy: First, set the batch size of the output layer to 32, use the AdamW optimizer, and set the initial learning rate to 2×10. -5 A linear learning rate decay strategy with warmup is adopted, with the weight decay coefficient set to 0.01 for regularization, the dropout rate set to 0.1, and early stopping based on validation set performance during training. The patience period is set to 3 epochs. There are no restrictions on the selection of key parameters; they can be parameters obtained experimentally.

[0115] Step 106: Based on disaster information and disaster response information, generate a disaster push page and push the disaster push page to the associated user terminals for display.

[0116] In some embodiments, the executing entity may generate a disaster push page based on the disaster information and disaster response information, and push the disaster push page to associated user terminals for display. The associated user terminals may be users within a preset area that are connected to the executing entity via wired or wireless connection. The preset area may be a pre-defined administrative division.

[0117] Optionally, after step 106, based on the aforementioned disaster response information, the associated emergency response equipment is controlled to perform the corresponding disaster emergency response task.

[0118] In some embodiments, the aforementioned implementing entity may, based on the aforementioned disaster response information, control associated emergency response equipment to perform corresponding disaster emergency response tasks. The aforementioned emergency response equipment may be equipment designed to respond to a specific disaster emergency response task. For example, if the aforementioned disaster information could be a wildfire caused by high temperatures, and the aforementioned disaster emergency response task could be extinguishing a fire in an area, then the aforementioned emergency response equipment could be a water-spraying drone.

[0119] The above embodiments of this disclosure have the following beneficial effects: The disaster information push method based on a large language model, as described in some embodiments of this disclosure, accurately determines the disaster level, avoiding casualties or waste of emergency response resources. Specifically, the reason for inaccurate disaster level determination and the resulting casualties or waste of emergency response resources is that when using convolutional neural networks to classify meteorological environmental data to determine the disaster level, the convolutional neural network cannot effectively fuse multi-dimensional features to determine the disaster level for structured data, unstructured data, and geographic information, leading to inaccurate disaster level determination and resulting in casualties or waste of emergency response resources. Based on this, the disaster information push method based on a large language model, as described in some embodiments of this disclosure, firstly, acquires environmental text information. This allows obtaining meteorological text data from the meteorological monitoring center. Secondly, based on the aforementioned environmental text information, feature extraction processing is performed to generate multiple extracted features, resulting in an extracted feature group. This allows the extraction of data features from the text data. Then, for each extracted feature in the extracted feature group, the extracted feature is classified to generate at least one local feature and at least one global feature. This allows the distinction between global and local features. Next, the at least one local feature and the at least one global feature are preprocessed to generate preprocessed local and global features, resulting in local feature sets and global feature sets. This allows for the preprocessing of local and global features. Then, the local and global feature sets are input into a pre-trained large language model to obtain disaster information and corresponding disaster response information. Thus, by fusing and analyzing structured data, unstructured data, and geographic information from the large language model's environment, the disaster level corresponding to the meteorological environment can be accurately determined, thereby avoiding casualties or waste of emergency response resources.

[0120] Further reference Figure 2As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a disaster information push device based on a large language model. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this disaster information push device based on a large language model can be specifically applied to various electronic devices.

[0121] like Figure 2 As shown, some embodiments of the disaster information push device 200 based on a large language model include: a receiving unit 201, an enhancement unit 202, a noise suppression unit 203, a classification unit 204, an input unit 205, and an optimization and parameter tuning unit 206. The receiving unit 201 is configured to receive at least one output vector sent by the decoder included in the large language model, wherein the output vector in the at least one output vector corresponds to dimensional information; the enhancement unit 202 is configured to enhance the at least one output vector based on the feature fusion module to generate an enhanced output vector set; the noise suppression unit 203 is configured to perform noise suppression processing on each enhanced output vector in the enhanced output vector set through the feature fusion module to generate a suppressed enhanced output vector, which serves as a suppressed output feature to obtain a suppressed output feature set; the classification unit 204 is configured to classify each suppressed output feature in the suppressed output feature set based on the hierarchical classification module to generate a classified output feature set, wherein the classified output features in the classified output feature set correspond to classification information; the input unit 205 is configured to input the classified output feature set into the loss function module to obtain a feature loss value set; and the optimization and parameter tuning unit 206 is configured to perform optimization and parameter tuning processing on the large language model based on the feature loss value set.

[0122] It is understandable that the units and references recorded in the disaster information push device 200 based on the large language model are... Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method are also applicable to the disaster information push device 200 based on a large language model and the units contained therein, and will not be repeated here.

[0123] The following is for reference. Figure 3 This document illustrates a structural schematic of an electronic device 300 suitable for implementing some embodiments of the present disclosure. The electronic devices in some embodiments of the present disclosure may include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 3The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0124] like Figure 3 As shown, the electronic device 300 may include a processing unit 301 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0125] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0126] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 309, or installed from storage device 308, or installed from ROM 302. When the computer program is executed by processing device 301, it performs the functions defined in the methods of some embodiments of this disclosure.

[0127] It should be noted that, in some embodiments of this disclosure, the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0128] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0129] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: acquire environmental text information, wherein the environmental text information is text information representing environmental information received from a target terminal through a preset interface; perform feature extraction processing based on the environmental text information to generate multiple extracted features, resulting in an extracted feature group; for each extracted feature in the extracted feature group, classify the extracted feature to generate at least one local feature and at least one global feature; preprocess the at least one local feature and the at least one global feature to generate preprocessed local features and global features, resulting in a local feature group and a global feature group; input the local feature group and the global feature group into a pre-trained large language model to obtain disaster information and corresponding disaster response information; and generate a disaster push page based on the disaster information and the disaster response information, and push the disaster push page to the associated user terminal for display.

[0130] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0131] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0132] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including an acquisition unit, an execution unit, a classification unit, a preprocessing unit, an input unit, and a generation unit. The names of these units do not necessarily limit the specific unit; for example, an acquisition unit may also be described as a "unit for acquiring environmental text information."

[0133] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0134] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A disaster information push method based on a large language model, comprising: Acquire environmental text information, wherein the environmental text information is text information used to characterize environmental information received from the target terminal through a preset interface; Based on the environmental text information, feature extraction processing is performed to generate multiple extracted features, resulting in an extracted feature group; For each extracted feature in the extracted feature group, the extracted feature is classified to generate at least one local feature and at least one global feature; The at least one local feature and the at least one global feature are preprocessed to generate preprocessed local features and global features, resulting in a local feature group and a global feature group. The local feature group and the global feature group are input into a pre-trained large language model to obtain disaster information and disaster response information corresponding to the disaster information. The large language model includes an output layer, which includes a feature fusion module, which includes local feature units and global feature units. The output layer is trained through the following steps: Receive at least one output vector sent by the decoder included in the large language model, wherein the output vector in the at least one output vector corresponds to dimension information; Based on the feature fusion module, the at least one output vector is enhanced to generate an enhanced output vector set, including: The at least one output vector is input into the local feature unit included in the feature fusion module to obtain the local feature set corresponding to the output vector; A global feature set is generated based on the local feature set and the global feature units; For each local feature in the local feature set, a global feature corresponding to the local feature is selected from the global feature set, and the local feature and the global feature are fused to generate a fused feature as the enhanced output vector. This includes: upsampling the global feature to generate a sampled global feature; determining the number of feature channels between the global feature and the local feature; adjusting the number of channels corresponding to the sampled global feature in response to the determination that the number of feature channels between the sampled global feature and the local feature is inconsistent, so that the number of channels between the sampled global feature and the local feature is consistent; performing a first convolution on the local feature to generate a first convolutional local feature; adding the first convolutional local feature and the sampled global feature element-wise to generate a summed feature; performing a second convolution on the summed feature to generate a convolutional fused feature, thus obtaining the convolutional feature; and batch normalizing the convolutional feature to generate a batch-normalized convolutional feature as the enhanced output vector. Each generated enhanced output vector is defined as the enhanced output vector set; Based on the disaster information and the disaster response information, a disaster push page is generated, and the disaster push page is pushed to the associated user terminal for display.

2. The method according to claim 1, wherein, The large language model includes: an input processing layer, an encoder layer, and a decoder layer; and The method further includes: Based on the disaster response information, the associated emergency response equipment is controlled to perform the corresponding disaster emergency response tasks.

3. The method according to claim 2, wherein, The output layer further includes: a hierarchical classification module and a loss function module; The output layer is also trained through the following steps: The feature fusion module performs noise suppression processing on each enhanced output vector in the enhanced output vector set to generate a suppressed enhanced output vector, which serves as the suppressed output feature to obtain the suppressed output feature set. Based on the hierarchical classification module, each suppressed output feature in the suppressed output feature set is classified to generate a classified output feature set, wherein the classified output features in the classified output feature set correspond to classification information. The classified output feature set is input into the loss function module to obtain the feature loss value set; Based on the aforementioned feature loss value set, the large language model is optimized and its parameters are tuned.

4. The method according to claim 3, wherein, The loss function module includes: main classification loss function, fine-grained classification loss function, and feature consistency loss function.

5. The method according to claim 3, wherein, The feature fusion module performs noise suppression processing on each enhanced output vector in the enhanced output vector set to generate a suppressed enhanced output vector, which serves as the suppressed output feature, thus obtaining a suppressed output feature set, including: For each enhanced output vector in the enhanced output vector set, the following suppression step is performed: Based on the gating unit included in the feature fusion module, the feature importance value corresponding to the enhanced output vector is determined; Based on the feature importance value, the enhanced output vector is subjected to feature filtering processing to generate a filtered output vector; The enhanced output vector and the filtered output vector are combined to generate a combined vector, which serves as the suppressed output feature.

6. The method according to claim 3, wherein, The hierarchical classification module includes a master classifier and a fine-grained classifier; and The step of classifying each suppressed output feature in the suppressed output feature set based on the hierarchical classification module to generate a classified output feature set includes: For each suppressed output feature in the set of suppressed output features, the following processing steps are performed: The suppressed output features are then subjected to dimensionality reduction processing to generate dimensionality-reduced features; The dimensionality reduction features are mapped to a preset category space to generate mapped features; Based on the fine-grained classifier included in the hierarchical classification module, the mapping features are subjected to fine-grained classification processing to generate fine-grained classified mapping features, which are used as the output features after classification.

7. A disaster information push device based on a large language model, comprising: The acquisition unit is configured to acquire environmental text information, wherein the environmental text information is text information used to characterize environmental information received from the target terminal through a preset interface; The execution unit is configured to perform feature extraction processing based on the environmental text information to generate multiple extracted features and obtain an extracted feature group; A classification unit is configured to classify each extracted feature in the extracted feature group to generate at least one local feature and at least one global feature. The preprocessing unit is configured to preprocess the at least one local feature and the at least one global feature to generate preprocessed local features and global features, resulting in a local feature group and a global feature group. An input unit is configured to input the local feature set and the global feature set into a pre-trained large language model to obtain disaster information and disaster response information corresponding to the disaster information. The large language model includes an output layer, which includes a feature fusion module, which includes local feature units and global feature units. The output layer is trained through the following steps: Receive at least one output vector sent by the decoder included in the large language model, wherein the output vector in the at least one output vector corresponds to dimension information; Based on the feature fusion module, the at least one output vector is enhanced to generate an enhanced output vector set, including: The at least one output vector is input into the local feature unit included in the feature fusion module to obtain the local feature set corresponding to the output vector; A global feature set is generated based on the local feature set and the global feature units; For each local feature in the local feature set, a global feature corresponding to the local feature is selected from the global feature set, and the local feature and the global feature are fused to generate a fused feature as the enhanced output vector. This includes: upsampling the global feature to generate a sampled global feature; determining the number of feature channels between the global feature and the local feature; adjusting the number of channels corresponding to the sampled global feature in response to the determination that the number of feature channels between the sampled global feature and the local feature is inconsistent, so that the number of channels between the sampled global feature and the local feature is consistent; performing a first convolution on the local feature to generate a first convolutional local feature; adding the first convolutional local feature and the sampled global feature element-wise to generate a summed feature; performing a second convolution on the summed feature to generate a convolutional fused feature, thus obtaining the convolutional feature; and batch normalizing the convolutional feature to generate a batch-normalized convolutional feature as the enhanced output vector. Each generated enhanced output vector is defined as the enhanced output vector set; The generation unit is configured to generate a disaster push page based on the disaster information and the disaster response information, and to push the disaster push page to the associated user terminal for display.

8. An electronic device, comprising: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 6.

9. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images