MAML-based cross-border public opinion analysis model, method, device and storage medium

Abstract

The application discloses a cross-border public opinion analysis model, method and device based on MAML and a storage medium. The pre-training base model of the cross-border public opinion analysis model adopts a double-channel architecture, performs multilingual semantic recognition on cross-border public opinion data, obtains a text embedding vector, and then performs gate fusion processing on cultural features corresponding to the region where the cross-border public opinion data is located and the text embedding vector. The fusion feature vector is subjected to public opinion sentiment classification to obtain a public opinion sentiment classification result. The cross-border public opinion analysis model optimizes model parameters, jointly learns meta-tasks defined by a task distribution, and the meta-tasks include intra-language tasks, cross-language tasks and cultural confrontation tasks. The objective function of the cross-border public opinion analysis model includes multi-task loss, cultural feature loss and semantic reconstruction loss. The embodiment of the application can effectively identify the sentiment tendency of different public opinion subjects in a cross-cultural context, and solves the problem of semantic deviation of a traditional public opinion analysis model in a cross-cultural context.

Description

Technical Field

[0001] This invention relates to the field of meta-learning technology, and in particular to a cross-border public opinion analysis model, method, device and storage medium based on MAML. Background Technology

[0002] Traditional sentiment classification techniques mainly consist of rule-based methods and machine learning-based methods. Rule-based methods primarily rely on experience or expert opinions to obtain statistical features and sentiment dictionaries to classify text for sentiment. Machine learning-based methods first establish a training set by manually labeling a portion of the data, then extract and learn features from the training set to build a classification model, and finally use the classification model to classify the sentiment of data with unknown labels.

[0003] However, with the surge in data volume and the increasing complexity of text information features, the shortcomings of the above two methods have gradually become apparent, as they rely too heavily on complex feature engineering. The emergence of deep learning has provided new means for natural language processing tasks. For example, Socher et al. proposed the RNTN (recursive neural tensor network) model based on recurrent neural networks for sentiment binary classification; Wang et al. used long short-term memory networks to analyze the sentiment polarity of text; Jiang et al. proposed the MANs (mutual attention neural networks) model for aspect-level sentiment classification; Zhao et al.

[15] modeled sentiment dependencies based on GCN (graph convolutional networks) and then performed sentiment classification. Deep learning models can automatically learn good features for the target problem and have achieved good results in text sentiment binary classification tasks, but there is still room for improvement in multi-classification tasks.

[0004] With globalization and the development of social media, the dissemination of public opinion regarding emergencies exhibits complex characteristics across cultures, platforms, and subjects. Traditional sentiment recognition models typically face two bottlenecks: first, multilingual semantic drift leads to a significant decline in the performance of cross-context models; second, existing methods struggle to reveal the potential cultural modulation and agenda-playing mechanisms within the "public-media-government" triadic public opinion structure.

[0005] Therefore, there is an urgent need to invent a model and method for cross-border public opinion analysis to solve the problems of semantic bias and low classification accuracy of existing public opinion sentiment classification methods in cross-cultural contexts. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a cross-border public opinion analysis model, method, device and storage medium based on MAML, which at least partially solves the problems existing in the prior art.

[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0009] According to a first aspect of the present invention, a cross-border public opinion analysis model based on MAML is provided, wherein the pre-trained base model of the cross-border public opinion analysis model adopts a dual-channel architecture;

[0010] The pre-trained basic model is used to perform multilingual semantic recognition on cross-border public opinion data to obtain the text embedding vector corresponding to the cross-border public opinion data. Then, the cultural features corresponding to the region where the cross-border public opinion data is located are subjected to gating fusion processing with the text embedding vector to obtain the fused feature vector. Finally, the fused feature vector is used for public opinion sentiment classification to obtain the public opinion sentiment classification result.

[0011] The cross-border public opinion analysis model optimizes model parameters. Joint learning is based on task distribution The defined meta-tasks include intra-language tasks, cross-language tasks, and cultural adversarial tasks.

[0012] The objective function of the cross-border public opinion analysis model includes multi-task loss, cultural feature loss, and semantic reconstruction loss.

[0013] The objective function of the cross-border public opinion analysis model The formula is: ,in, For multi-tasking loss, For the loss of cultural characteristics, For semantic reconstruction loss based on the variational autoencoder framework, , and For the loss term coefficient;

[0014] The The formula for calculating the multi-task loss is: , ,in, For cross-cultural sensitive word database, This is an empirical coefficient. To adjust the weights, For indicator functions, This represents the true label of the i-th sample. This represents the i-th input sample. The parameter is The model for input The predicted output;

[0015] The loss of cultural characteristics The calculation formula is: , ,in, Let S be the culture-adaptation vector, and S be the culture similarity matrix between samples within a batch. This is the preset temperature coefficient.

[0016] Furthermore, the pre-trained base model adopts the XLM-RoBERTa-large model, and a contrastive learning loss term is introduced in the semantic recognition stage. , ,in, Embedding for language identifiers, For statement embedding, Preset boundary threshold;

[0017] If the difference between the cosine similarity of the language identifier embedding and the cosine similarity of the corresponding sentence embedding exceeds the preset boundary threshold, the gradient penalty mechanism of the pre-trained base model is triggered. While preserving language features, the semantically equivalent representations across languages ​​are mapped to neighboring vector regions to obtain a unified multilingual joint representation space.

[0018] Furthermore, the pre-trained base model includes a cultural feature encoder;

[0019] The cultural feature encoder is used to generate a cultural adaptation vector based on a fully connected network, according to the cultural features corresponding to the region where the cross-border public opinion data is located. The cultural characteristics include Hofstede cultural multidimensional indicators and Globe cultural clustering characteristics. The Hofstede cultural multidimensional indicators include power distance (PDI), individualism (IDV), uncertainty avoidance (UAI), long-term orientation (LTO), and permissiveness (IND).

[0020] The cultural feature encoder is further used to perform gated fusion processing on the cultural adaptation vector and the text embedding vector to obtain a fused feature vector. The fusion formula is as follows:

[0021]

[0022]

[0023] in, To fuse feature vectors, For the Sigmoid function, These are learnable parameters used to dynamically weight the culture adaptation vector and the text embedding vector.

[0024] Furthermore, the intralingual task is used for sentiment classification in different cultural regions of the same language, the cross-lingual task is used for sentiment classification on parallel corpora of different languages, and the cultural adversarial task is used to suppress culture-related features through a gradient inversion layer.

[0025] Furthermore, the cross-border public opinion analysis model adopts an internal and external dual-loop gradient update mechanism when dealing with new cultural scenarios;

[0026] In the inner loop phase, the task loss gradient is calculated using the support set data to make initial adjustments to the model parameters. The formula is as follows: ,in, , To preset the base learning rate, This refers to cultural distance measured using Mahalanobis distance with Hofstede's cultural dimension. This is a cultural difference coefficient used to dynamically adjust the learning rate. This represents the i-th specific task. Indicates in the task loss function on, The parameter is The model function;

[0027] In the outer loop phase, the query set data is used to aggregate multi-task loss optimization meta-parameters, with the following formula: ,in, The learning rate is the outer loop rate.

[0028] Furthermore, the cross-border public opinion analysis model adopts an adaptive meta-parameter update strategy, using a cultural gradient pruning algorithm to calculate the angle between the parameter update direction and the cultural principal component. If the angle is greater than a preset angle, the gradient is scaled proportionally to avoid distortion of the cultural feature space.

[0029] The cross-border public opinion analysis model dynamically adjusts the outer loop learning rate. The formula is: ,in, To preset the maximum learning rate, To preset the minimum learning rate, This is the preset total number of iterations.

[0030] According to a second aspect of the present invention, a cross-border public opinion analysis method based on MAML is provided, the method comprising:

[0031] Cross-border public opinion data is input into a cross-border public opinion analysis model based on MAML as described in any of the above items to obtain public opinion sentiment classification results.

[0032] According to a third aspect of the present invention, a cross-border public opinion analysis device based on MAML is provided, the device comprising: a processor and a memory;

[0033] The memory is used to store one or more program instructions;

[0034] The processor is configured to run one or more program instructions to perform the steps of a MAML-based cross-border public opinion analysis method as described in any of the preceding claims.

[0035] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, wherein a cross-border public opinion analysis model based on MAML as described in any of the preceding claims is stored on the computer-readable storage medium.

[0036] This invention provides a MAML-based cross-border public opinion analysis model, method, device, and storage medium. The pre-trained base model of the cross-border public opinion analysis model adopts a dual-channel architecture, performing multilingual semantic recognition on cross-border public opinion data to obtain text embedding vectors. Then, it performs gating fusion processing on the cultural features corresponding to the region where the cross-border public opinion data is located and the text embedding vectors, and performs public opinion sentiment classification on the fused feature vectors to obtain the public opinion sentiment classification results. The cross-border public opinion analysis model optimizes model parameters and jointly learns meta-tasks defined by task distributions, including intra-lingual tasks, cross-lingual tasks, and cultural adversarial tasks. The objective function of the cross-border public opinion analysis model includes multi-task loss, cultural feature loss, and semantic reconstruction loss. This invention can effectively identify the sentiment tendencies of different public opinion subjects in cross-cultural contexts, solving the semantic bias problem of traditional public opinion analysis models in cross-cultural contexts. Attached Figure Description

[0037] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the architecture of a cross-border public opinion analysis model based on MAML provided in an embodiment of the present invention. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0040] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0041] Figure 1 This diagram illustrates the architecture of a cross-border public opinion analysis model based on MAML provided by an embodiment of the present invention.

[0042] The cross-border public opinion analysis model based on MAML provided in this embodiment of the invention adopts a transfer learning framework based on an improved Model-Agnostic Meta-Learning (MAML).

[0043] The pre-trained base model XLM-RoBERTa-large of the aforementioned cross-border public opinion analysis model adopts a dual-channel architecture and has the ability to jointly represent multiple languages.

[0044] The aforementioned pre-trained basic model introduces a contrastive learning loss term in the semantic recognition stage. , ,in, Embedding for language identifiers, For statement embedding, This is a preset boundary threshold.

[0045] If the difference between the cosine similarity of the language identifier embedding and the cosine similarity of the corresponding sentence embedding exceeds a preset boundary threshold, the gradient penalty mechanism of the pre-trained base model is triggered. While preserving language features, the semantically equivalent representations across languages ​​are mapped to neighboring vector regions, ultimately resulting in a unified multilingual joint representation space.

[0046] The aforementioned pre-trained base model also includes a cultural feature encoder. The input to the aforementioned cultural feature encoder is the cultural features corresponding to the region where the cross-border public opinion data is located. These cultural features include Hofstede cultural multidimensional indicators and Globe cultural clustering features. The aforementioned Hofstede cultural multidimensional indicators include Power Distance (PDI), Individualism (IDV), Uncertainty Avoidance (UAI), Long-Term Orientation (LTO), and Permissiveness (IND).

[0047] Using fully connected networks, cultural adaptation vectors are generated based on cultural characteristics. .

[0048] The cultural feature encoder is also used to perform gated fusion of the cultural adaptation vector and the text embedding vector to obtain a fused feature vector. The fusion formula is as follows:

[0049]

[0050]

[0051] in, To fuse feature vectors, For the Sigmoid function, These are learnable parameters used to dynamically weight the culture adaptation vector and the text embedding vector.

[0052] The meta-learning strategy of the cross-border public opinion analysis model is: to optimize model parameters. Joint learning is based on task distribution The defined meta-tasks include intra-lingual tasks, cross-lingual tasks, and cultural adversarial tasks. The intra-lingual tasks are used for sentiment classification in different cultural regions of the same language, the cross-lingual tasks are used for sentiment classification of parallel corpora in different languages, and the cultural adversarial tasks are used to suppress culture-related features through gradient inversion layers (GRL).

[0053] To address the modulating effect of cultural background on emotional expression, a multi-granularity loss function is designed for the cross-border public opinion analysis model. The objective function of the cross-border public opinion analysis model includes multi-task loss, cultural feature loss, and semantic reconstruction loss, as shown in the formula: ,in, For multi-tasking loss, For the loss of cultural characteristics, For semantic reconstruction loss based on the variational autoencoder framework, , and This is the coefficient for the loss term.

[0054] Specifically, The formula for calculating the multi-task loss is: , ,in, It is a cross-cultural sensitive word database (covering political metaphors, religious taboos, etc.). This is an empirical coefficient. To adjust the weights, For indicator functions, This represents the true label of the i-th sample. This represents the i-th input sample. The parameter is The model for input The predicted output.

[0055] Loss of cultural characteristics The calculation formula is: , ,in, Let S be the culture adaptation vector, and S be the culture similarity matrix between samples within the batch. , The preset temperature coefficient is (preferably 0.07).

[0056] The embodiments of the present invention significantly reduce the phenomenon of emotional drift by introducing a cross-semantic boundary adjustment factor and a cultural semantic bias penalty term.

[0057] When dealing with new cultural scenarios, the cross-border public opinion analysis model adopts an internal and external dual-loop gradient update mechanism.

[0058] In the inner loop phase, the task loss gradient is calculated using the support set data to make initial adjustments to the model parameters. The formula is as follows: ,in, , To preset the base learning rate, , This refers to cultural distance measured using Mahalanobis distance with Hofstede's cultural dimension. This is a cultural difference coefficient used to dynamically adjust the learning rate. This represents the i-th specific task. Indicates in the task loss function on, The parameter is The model function.

[0059] In the outer loop phase, the query set data is used to aggregate multi-task loss optimization meta-parameters, with the following formula: ,in, The learning rate is the outer loop rate.

[0060] The embodiments of the present invention achieve culturally sensitive adaptive learning rate adjustment through an internal and external dual-loop gradient update mechanism, forming a meta-learning optimization system for cross-cultural scenarios.

[0061] Traditional MAML is prone to getting trapped in local optima in cross-cultural scenarios. This invention addresses this by proposing an adaptive meta-parameter update strategy. The cross-border public opinion analysis model utilizes a cultural gradient pruning algorithm to calculate the angle between the parameter update direction and the cultural principal components. If the included angle is greater than the preset angle To avoid spatial distortion of cultural features, the gradient is scaled proportionally, using the following formula: .

[0062] The cross-border public opinion analysis model uses a cosine annealing program to dynamically adjust the outer loop learning rate. The formula is: ,in, To preset the maximum learning rate, , To preset the minimum learning rate, Preset total number of iterations .

[0063] This invention selects the 7.9 magnitude earthquake in Country A in 2025 as a sample for cross-border public opinion analysis. First, the geopolitical sensitivity drives the typicality of the sample, as Country A has long been plagued by multiple geopolitical conflicts. The earthquake not only triggered a natural disaster response but also exacerbated deeper contradictions such as international aid and sovereignty disputes, providing a typical scenario for observing the interplay of multiple actors in cross-border public opinion. Second, the complexity of multi-agent participation meets the research needs of this invention. In this event, the local public in Country A expressed their doubts and approvals through social media (such as MySQUAR); international media dominated global public opinion; and the government of Country A and international organizations competed for agenda dominance in aid efforts. This three-pronged interactive pattern of "public-media-government" fully covers the core actors in cross-border public opinion dissemination and can systematically reveal the dynamic patterns of multi-agent collaboration and confrontation. Regarding data representativeness, the study uses the earthquake's occurrence time as a benchmark, collecting data from before the earthquake to 30 days after (a total of 1.2 million data points), covering social media within Country A, regional media, and global dissemination nodes. By using stratified sampling and geotagging techniques, we ensure a balance between temporal continuity, spatial coverage, and subject diversity in the data, avoiding data bias from a single platform or language.

[0064] Cross-border public opinion data exhibits a deep interweaving of multi-source heterogeneity and subject heterogeneity, specifically manifested in significant differences in information production among different subjects and dynamic differentiation in spatiotemporal evolution. Within country A, public social media exhibits emotional polarization based on a binary narrative of "disaster relief demands - national sentiment," with negative emotions often intertwined with memories of historical conflicts. For example, posts on the topic of delayed rescue within 72 hours after the earthquake contained metaphors for the Rakhine conflict. In the spatiotemporal dimension, public opinion evolution shows a four-stage differentiation: from the geographically concentrated dissemination of pre-earthquake warning information, to the differentiation of topics across cross-border platforms within 72 hours after the disaster, then to the geopolitical game period, and finally forming a local ripple effect in the post-disaster reconstruction narrative.

[0065] Emotional dissemination in such geopolitically sensitive events is modulated by both historical memory and international discourse, while the agenda-building game between international organizations and sovereign states exhibits a window effect—the dominant power in communication shifts dynamically with the event cycle. These characteristics reveal that cross-border public opinion analysis needs to construct multimodal data processing algorithms, integrate dynamic weight adjustment mechanisms and cross-cultural symbolic knowledge bases, and highlight the methodological value of a multi-stakeholder framework in decoding complex public opinion ecosystems.

[0066] Using the 7.9 magnitude earthquake in Country A in 2025 as a sample of cross-border public opinion, this invention proposes a cross-border public opinion analysis model based on MAML. In addition to the conventional accuracy and F1 score, three cross-cultural evaluation indicators are added: cultural semantic deviation rate (CSD), cross-linguistic stability (CLS), and metaphor recognition rate (MRR).

[0067] Specifically, the Cultural Semantic Deviation Rate (CSD) is used to quantify the model's semantic understanding bias towards culturally sensitive words, and the calculation formula is as follows: ,in, This represents a set of culturally sensitive words annotated by experts. This is an indicator function.

[0068] When sample belong When the model predicts With real labels Inconsistencies are counted as 1. The lower the Cultural Semantic Deviation Rate (CSD) value, the stronger the model's cultural semantic alignment ability.

[0069] Cross-linguistic stability (CLS) is used to evaluate the predictive consistency of a model in a bilingual scenario. Its calculation formula is as follows: ,in, This represents the set of samples that correctly predict the language in the k-th language. CLS = 1 when the predictions in both languages ​​are completely consistent; if there is a discrepancy in the predictions between languages, the CLS value will be greater than 1. The closer the CLS value is to 1, the higher the stability of the model's cross-language transfer.

[0070] Metaphor Recognition Rate (MRR) is used to evaluate the accuracy of the sentiment mapping of political metaphors using an expert-annotated test set. The sentiment polarity (positive / negative / neutral) of political metaphor statements is independently annotated by linguistic experts. The higher the MRR value, the stronger the model's ability to analyze the deep semantics of cultural metaphors.

[0071] The embodiments of this invention set up five sets of control experiments to verify the improvement effect of the model algorithm: Baseline 1 is the Chinese sentiment dictionary method based on SnowNLP, Baseline 2 is the monolingual fine-tuning BERT (independent training of language A / Chinese / English), Baseline 3 is the cross-language model XLM-Twitter, Baseline 4 is the traditional MAML framework, and Proposed is the improved cross-border public opinion analysis model based on MAML provided in the embodiments of this invention.

[0072] The results of the comparative experiment are shown in Table 1:

[0073] Table 1 Comparative test results

[0074]

[0075] As can be seen, in terms of accuracy, Proposed achieved an accuracy of 82.3%, which is 21.0 and 5.1 percentage points higher than Baseline1 (61.3%) and Baseline4 (77.2%), respectively; Proposed's F1 score reached 79.6%, which is 5.8 percentage points higher than the best baseline Baseline4 (73.8%).

[0076] Of particular note is the Cultural Semantic Difference (CSD) index, which decreased from 29.4% in Baseline 1 to 9.2% in Proposed, a drop of 20.2 percentage points, verifying the effectiveness of the cultural adaptation mechanism provided by the embodiments of the present invention.

[0077] The Proposed also performed exceptionally well in terms of Cross-Cultural Learning Stability (CLS) and Cross-Cultural Sensitive Word Recognition Accuracy (MRR), setting new benchmarks at 83.7% and 91.2% respectively, representing improvements of 9.5 and 22.8 percentage points over Baseline4.

[0078] The gradient changes in the experimental data show that as the depth of model improvement increases, the CSD index shows a monotonically decreasing trend, while CLS and MRR show a significant positive correlation growth, indicating that the multi-granularity cultural transfer strategy proposed in this invention has a systematic advantage in reducing semantic bias and improving cultural sensitivity.

[0079] The effects of module optimization were compared through ablation experiments. The results of the ablation experiments are shown in Table 2.

[0080] Table 2 Ablation Test Table

[0081]

[0082] It is evident that removing the "cultural adaptation channel (dual-channel architecture)" leads to a significant decline in core metrics, with the F1 score dropping to 9.7%, the cultural semantic difference (CSD) increasing by 12.3 percentage points, and the cross-cultural sensitive word recognition accuracy (MRR) falling to 25.6%. Removing the "dynamic loss function" further deteriorates the F1 score (6.3%) and MRR (18.4%), and the increase in CSD narrows to 7.8%. Removing the "meta-learning rate annealing strategy" results in the most drastic performance collapse, with the F1 score remaining at only 4.1%. At this point, the CSD only increases by 5.2%, but the MRR plummets to 9.7%.

[0083] The three sets of comparative experiments above show that the "cultural adaptation channel" contributes the most to CSD regulation (responsible for offsetting 70.7% of potential semantic bias), the "dynamic loss function" dominates the improvement of F1 score (providing 82.4% of basic performance gain), and the "meta-learning rate annealing strategy" is the core factor in maintaining the MRR index (contributing 89.4%).

[0084] In summary, this invention provides a systematic cross-border online public opinion analysis model for the complex cross-border and cross-cultural public opinion dissemination phenomenon involving multiple stakeholders in public emergencies.

[0085] Unlike traditional single-subject research paradigms, this paper integrates the perspectives of three major subjects: the public, the media, and domestic and foreign governments. By collecting multi-dimensional spatiotemporal public opinion data before and after the earthquake (covering heterogeneous data from multiple sources such as social media, news portals, and government announcements), and combining life cycle theory and complex network analysis methods, it deeply deconstructs the spatiotemporal differentiation characteristics of public opinion dissemination under special geographical conditions.

[0086] The improved MAML-based cross-border sentiment analysis model proposed in this invention can effectively identify the sentiment biases of different subjects in cross-cultural contexts. Experimental data shows that in international emergencies, commonly used traditional sentiment analysis models suffer from cross-cultural semantic bias.

[0087] In addition, this invention also provides a cross-border public opinion analysis method based on MAML, which includes: inputting cross-border public opinion data into a cross-border public opinion analysis model based on MAML as described above, and obtaining public opinion sentiment classification results.

[0088] In addition, this invention also provides a cross-border public opinion analysis device based on MAML, the device comprising: a processor and a memory; the memory for storing one or more program instructions; the processor for running one or more program instructions to execute the steps of a cross-border public opinion analysis method based on MAML as described above.

[0089] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a cross-border public opinion analysis model based on MAML as described above.

[0090] In summary, this invention takes the cross-border online public opinion surrounding the 7.9 magnitude earthquake in Country A in 2025 as its research object, and proposes a cross-border public opinion analysis model based on MAML. This model solves the semantic bias problem of traditional public opinion analysis models in cross-cultural contexts. The main technical effects are as follows:

[0091] First, the improved MAML algorithm constructed in this invention effectively overcomes the semantic barriers in cross-cultural sentiment analysis. Experimental results show that, through a dual-channel meta-learning architecture and a cultural feature fusion mechanism, the model achieves a sentiment classification accuracy of 82.3% in cross-border online public opinion regarding the 7.9 magnitude earthquake in Country A, an improvement of 6.9 percentage points compared to traditional cross-language models. Particularly in handling culturally sensitive words, the cultural semantic bias rate (CSD) decreased from 29.4% in the baseline model to 9.2%, while the metaphor recognition rate (MRR) increased to 91.2%. This indicates that the parameter transfer strategy based on cultural dimensions can effectively capture the deep semantic features in the context of Country A, providing a new methodological framework for cross-cultural semantic understanding in the field of digital governance.

[0092] Second, this invention reveals a three-fold differentiation pattern in public opinion dissemination under specific geopolitical environments. Spatiotemporal analysis shows that public opinion regarding the earthquake in Country A exhibits a "four-stage, three-concentric" dissemination pattern: within 72 hours of the earthquake, an "emotional storm eye" centered on questioning the effectiveness of disaster relief efforts formed on Country A's social media; in the second week, geopolitical issues triggered an "agenda-tearing" effect in regional media; and finally, during the post-disaster reconstruction period, a "core-regional-global" ripple diffusion structure was formed. Subject-based analysis shows a significant asymmetry in the "media-public" emotional transmission chain; the influence coefficient of the international media's humanitarian framework on public sentiment is 2.1 times that of government statements, but the trust threshold of the domestic public in Country A for government disaster relief data is significantly higher than their response to international media. This finding provides a key time window and path selection basis for cross-border public opinion intervention.

[0093] Third, this invention reveals a coupling mechanism of "cultural modulation and agenda-playing." Through an improved complex network analysis model, this invention quantifies the moderating effect of cultural cognitive differences on public opinion dissemination: the public's acceptance of international aid in Country A is negatively correlated with the power distance index in its cultural dimension, but positively correlated with the long-term orientation index. Simultaneously, the agenda-playing between the government and international organizations exhibits a "dual-cycle oscillation" characteristic: during the disaster emergency period (0-72 hours post-earthquake), the government of Country A achieves agenda dominance through data visualization strategies; while during the reconstruction period (15-30 days), the coordination framework of regional organizations gradually takes a dominant position. This dynamic game mechanism provides a decision-making basis for constructing a multi-stakeholder collaborative governance system.

[0094] This invention possesses dual innovative value in the field of digital governance: Methodologically, the proposed cross-cultural MAML framework achieves a balance between "interpretability of cultural features" and "unity of semantic representation" through a meta-learning mechanism, providing a new paradigm for multilingual public opinion analysis; in terms of governance practice, the constructed "three-dimensional, four-order" analysis model reveals the patterns of public opinion linkage in emergencies, and the proposed key parameters, such as the cultural adaptation correction coefficient and the agenda-playing window period (0-72 hours), can be directly applied to the optimization design of cross-border public opinion monitoring systems. The theoretical framework and methodological system formed by the embodiments of this invention can provide new analytical tools and practical paths for risk communication and public opinion guidance in global governance in the digital age.

[0095] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in this embodiment of the invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in this embodiment of the invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods. The storage medium can be memory, for example, volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM).The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory. Those skilled in the art will recognize that the functions described in the above examples can be implemented using a combination of hardware and software. When applied software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium accessible to general-purpose or special-purpose computers. Although the invention has been described in detail above with general description and specific embodiments, modifications or improvements can be made to it, which will be apparent to those skilled in the art. Therefore, such modifications or improvements made without departing from the spirit of the invention are all within the scope of protection claimed by the invention.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, or alterations made by those skilled in the art using the disclosed technical content shall fall within the protection scope of the present invention.

Claims

1. A cross-border public opinion analysis model based on MAML, characterized in that, The pre-trained base model of the cross-border public opinion analysis model adopts a dual-channel architecture; The pre-trained basic model is used to perform multilingual semantic recognition on cross-border public opinion data to obtain the text embedding vector corresponding to the cross-border public opinion data. Then, the cultural features corresponding to the region where the cross-border public opinion data is located are subjected to gating fusion processing with the text embedding vector to obtain the fused feature vector. Finally, the fused feature vector is used for public opinion sentiment classification to obtain the public opinion sentiment classification result. The cross-border public opinion analysis model optimizes model parameters. Joint learning is based on task distribution The defined meta-tasks include intra-language tasks, cross-language tasks, and cultural adversarial tasks. The objective function of the cross-border public opinion analysis model includes multi-task loss, cultural feature loss, and semantic reconstruction loss. The objective function of the cross-border public opinion analysis model The formula is: ,in, For multi-tasking loss, For the loss of cultural characteristics, For semantic reconstruction loss based on the variational autoencoder framework, , and For the loss term coefficient; The The formula for calculating the multi-task loss is: , ,in, For cross-cultural sensitive word database, This is an empirical coefficient. To adjust the weights, For indicator functions, This represents the true label of the i-th sample. This represents the i-th input sample. The parameter is The model for input The predicted output; The loss of cultural characteristics The calculation formula is: , ,in, Let S be the culture-adaptation vector, and S be the culture similarity matrix between samples within a batch. This is the preset temperature coefficient.

2. The cross-border public opinion analysis model based on MAML according to claim 1, characterized in that, The pre-trained base model adopts the XLM-RoBERTa-large model, and a contrastive learning loss term is introduced in the semantic recognition stage. , ,in, Embedding for language identifiers, For statement embedding, Preset boundary threshold; If the difference between the cosine similarity of the language identifier embedding and the cosine similarity of the corresponding sentence embedding exceeds the preset boundary threshold, the gradient penalty mechanism of the pre-trained base model is triggered. While preserving language features, the semantically equivalent representations across languages ​​are mapped to neighboring vector regions to obtain a unified multilingual joint representation space.

3. The cross-border public opinion analysis model based on MAML according to claim 2, characterized in that, The pre-trained base model includes a cultural feature encoder; The cultural feature encoder is used to generate a cultural adaptation vector based on a fully connected network, according to the cultural features corresponding to the region where the cross-border public opinion data is located. The cultural characteristics include Hofstede cultural multidimensional indicators and Globe cultural clustering characteristics. The Hofstede cultural multidimensional indicators include power distance (PDI), individualism (IDV), uncertainty avoidance (UAI), long-term orientation (LTO), and permissiveness (IND). The cultural feature encoder is further used to perform gated fusion processing on the cultural adaptation vector and the text embedding vector to obtain a fused feature vector. The fusion formula is as follows: in, To fuse feature vectors, For the Sigmoid function, These are learnable parameters used to dynamically weight the culture adaptation vector and the text embedding vector.

4. The cross-border public opinion analysis model based on MAML according to claim 1, characterized in that, The intralingual task is used for sentiment classification in different cultural regions of the same language, the cross-lingual task is used for sentiment classification on parallel corpora of different languages, and the cultural adversarial task is used to suppress culture-related features through a gradient inversion layer.

5. The cross-border public opinion analysis model based on MAML according to claim 1, characterized in that, The cross-border public opinion analysis model adopts an internal and external dual-loop gradient update mechanism when dealing with new cultural scenarios. In the inner loop phase, the task loss gradient is calculated using the support set data to make initial adjustments to the model parameters. The formula is as follows: ,in, , To preset the base learning rate, This refers to cultural distance measured using Mahalanobis distance with Hofstede's cultural dimension. This is a cultural difference coefficient used to dynamically adjust the learning rate. This represents the i-th specific task. Indicates in the task loss function on, The parameter is The model function; In the outer loop phase, the meta-parameters for multi-task loss optimization are aggregated using query set data, with the following formula: ,in, The learning rate is the outer loop rate.

6. The cross-border public opinion analysis model based on MAML according to claim 5, characterized in that, The cross-border public opinion analysis model adopts an adaptive meta-parameter update strategy and uses a cultural gradient pruning algorithm to calculate the angle between the parameter update direction and the cultural principal component. If the angle is greater than the preset angle, the gradient is scaled proportionally to avoid distortion of the cultural feature space. The cross-border public opinion analysis model dynamically adjusts the outer loop learning rate. The formula is: ,in, To preset the maximum learning rate, To preset the minimum learning rate, This is the preset total number of iterations.

7. A cross-border public opinion analysis method based on MAML, characterized in that, The method includes: Cross-border public opinion data is input into a cross-border public opinion analysis model based on MAML as described in any one of claims 1 to 6 to obtain public opinion sentiment classification results.

8. A cross-border public opinion analysis device based on MAML, characterized in that, The device includes: a processor and a memory; The memory is used to store one or more program instructions; The processor is configured to run one or more program instructions to perform the steps of the MAML-based cross-border public opinion analysis method as described in claim 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a cross-border public opinion analysis model based on MAML as described in any one of claims 1 to 6.

Images