International market public opinion analysis and business opinion mining system based on deep learning

Abstract

The invention discloses an international market public opinion analysis and business opportunity mining system based on deep learning, and relates to the technical field of cross-language emotion calculation, causal reasoning and anomaly detection. The system comprises a multi-source public opinion data acquisition module, a cross-cultural emotion recognition module, a causal graph neural network analysis module, an emotion anti-fact simulation module, a cross-context nested causal chain business opportunity discovery module, an abnormal emotion group detection module, a causal graph generation type learning engine and a business opportunity early warning output module. Through cooperative work of all the modules, the system can automatically identify hidden causal relationship chains from public opinion data of different languages in the world, predict potential market influences in advance, discover commercial opportunities such as supply gaps and sudden demand increase, and give an early warning for abnormal public opinions. According to the system provided by the invention, the depth and accuracy of international market public opinion analysis are remarkably improved, and interpretable business opinion intelligence and decision suggestions can be provided in time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, natural language processing and business intelligence, in particular to an international market public opinion analysis and business opportunity mining system based on deep learning. BACKGROUND

[0002] With the development of globalization and the Internet, the influence of public opinion information (such as social media posts, news reports, user reviews, etc.) in various markets on business decisions is increasing. However, due to language and cultural differences, the emotional expression and interpretation of the same event by people in different regions are significantly different. Existing public opinion analysis techniques mostly focus on a single language or cultural background, and lack effective means for cross-language emotion recognition and comparison, making it difficult to obtain a comprehensive and objective understanding of global public opinion trends. For example, most current multilingual event causal relationship identification methods mainly target English corpus and have lower performance in other languages, making it difficult to accurately identify event causal relationships in low-resource language environments. This limitation makes it difficult for traditional techniques to capture deep emotional biases in different market cultures, potentially missing important early warning signals.

[0003] In addition, the relationship between public opinion and market behavior is very complex, often not a simple correlation but a causal chain. For example, research has shown that there is a significant causal relationship between social media sentiment and stock market fluctuations, which can be used to predict market trends. However, existing methods are mostly limited to using sentiment indicators to predict financial market indices, and lack the ability to discover non-obvious business opportunities (such as emerging demand gaps and supply chain bottlenecks). Moreover, they usually react after the trend has become apparent, and cannot provide early insight into potential market impacts before a hot event erupts.

[0004] In terms of causal analysis, traditional machine learning methods mostly stay at the correlation level and are difficult to identify the causal path hidden in the evolution of public opinion. Especially in the complex global market environment, a chain reaction triggered by an event in a certain region or industry may cross language and cultural boundaries, forming a cross-context nested causal chain, and existing technologies lack effective modeling means for this.

[0005] At the same time, in the face of massive public opinion data, abnormal emotional groups (such as the sudden strong emotions of niche groups) are often early signals of significant business opportunities or risks. However, traditional public opinion monitoring is mostly based on statistical thresholds or manual rules, making it difficult to discover hidden abnormal patterns in a timely manner. In recent years, techniques such as generative adversarial networks (GAN) have been used for anomaly detection in text data to generate realistic samples and assist in identifying difficult-to-detect anomalies in the distribution. However, in the context of cross-cultural public opinion, there is no technology that uses GAN for the detection and simulation of emotional abnormal groups to improve the early warning capabilities of business opportunities.

[0006] Therefore, there is an urgent need for a new technical solution to overcome the above challenges and realize in-depth analysis of international market public opinion and business opportunity mining. SUMMARY

[0007] Technical purpose: In view of the shortcomings of the prior art, the present application discloses an international market public opinion analysis and business opportunity mining system based on deep learning, which can automatically mine potential business opportunities from global multilingual public opinion.

[0008] Technical solution: In order to achieve the above technical purpose, the present application adopts the following technical solution:

[0009] An international market public opinion analysis and business opportunity mining system based on deep learning, comprising a multi-source public opinion data acquisition module, a cross-cultural emotion recognition module, a causal graph neural network analysis module, an emotion counterfactual simulation module, a cross-context nested causal chain business opportunity discovery module, an abnormal emotion group detection module, a causal graph generative learning engine and a business opportunity early warning output module; wherein:

[0010] The multi-source public opinion data acquisition module is used to acquire public opinion text data from different countries and languages, and to acquire electroencephalogram signal data for cross-cultural calibration in the training stage;

[0011] The cross-cultural emotion recognition module performs sentiment analysis on public opinion texts of each language based on a deep learning model, and outputs emotion representations that can be compared across languages, which are trained by electroencephalogram signals to eliminate emotion recognition bias caused by different languages and cultures;

[0012] The causal graph neural network analysis module uses a multi-granularity causal graph neural network to construct a causal relationship graph of public opinion and market events, and identifies implicit causal paths in the evolution of public opinion at different time and semantic scales;

[0013] The emotion counterfactual simulation module uses the causal relationship graph to simulate the emotion evolution and market impact of target public opinion events in multiple scenarios, compares the differences in market-related indicators under different hypothetical scenarios, and is used to predict potential market changes;

[0014] The cross-context nested causal chain business opportunity discovery module extracts cross-language and cross-regional nested causal chains from the causal relationship graph and emotion simulation results, analyzes global supply or demand changes triggered by public opinion, identifies potential business opportunities and evaluates their impact and feasibility;

[0015] The abnormal emotion group detection module monitors abnormal emotion pattern groups in public opinion data based on a generative adversarial network model, generates an early warning signal when an emotion group deviating from the normal pattern is detected, and provides abnormal emotion pattern features to the cross-context nested causal chain business opportunity discovery module for further analysis;

[0016] The causal diagram generation learning engine, as a unified framework, receives information from the causal diagram neural network analysis module, the emotional counterfactual simulation module, the abnormal emotional group detection module, and the output feedback, continuously updates the generative adversarial network model and the causal relationship graph under unsupervised or weakly supervised conditions, so that the system gradually adapts to new public opinion dynamics.

[0017] The business opportunity early warning output module outputs the public opinion analysis report and the business opportunity mining result, including the causal chain, the trend prediction, and the corresponding decision suggestion or early warning prompt, based on the analysis results of the above modules.

[0018] Preferably, the cross-cultural emotion recognition module includes a multilingual text encoding submodule and an emotion prediction submodule; the output of the emotion prediction submodule is trained by aligning with the emotional features extracted from the electroencephalogram signal, specifically: for the training sample containing the electroencephalogram signal label, the error between the text emotion prediction value and the corresponding electroencephalogram emotion feature is minimized, so that the emotional representation of each language text output by the deep learning model is consistent with the actual emotional response of different cultural groups.

[0019] Preferably, the causal diagram neural network analysis module adopts a hierarchical graph neural network structure, processes the causal diagram nodes by granularity, and considers different time lags to identify short-term and long-term causal relationships; the causal diagram neural network analysis module learns the causal diagram structure through causal effect test and sparse regularization constraint to distinguish the true causal relationship and the correlation relationship only.

[0020] Preferably, the emotional counterfactual simulation module includes: a scenario setting unit for defining at least two different development scenarios; a simulation calculation unit for iteratively calculating the change trajectories of public opinion emotion indicators and related market indicators over time under each scenario based on the current causal relationship graph; and a result comparison unit for comparing the simulation results of different scenarios to evaluate the potential impact of a specific intervention or event on the market, outputting key difference indicators and recommended decision points.

[0021] Preferably, the cross-context nested causal chain business opportunity discovery module identifies the context attribute differences of different nodes in the causal chain to determine whether the chain crosses different languages, regions, or industries; for the cross-context causal chain, further evaluate its reliability and influence range, and output the high-reliability and high-influence chain as a potential business opportunity; the cross-context nested causal chain business opportunity discovery module also combines the scenario analysis of the emotional counterfactual simulation module to determine whether there is an intervention point that can significantly improve the business opportunity income by changing a certain antecedent scenario, and gives the corresponding action suggestion.

[0022] Preferably, the generative adversarial network of the abnormal sentiment group detection module comprises two parts: a discriminator and a generator. The discriminator learns the distribution of normal public opinion sentiment patterns and outputs the probability that actual public opinion data is normal or abnormal when input. The generator generates simulated public opinion data samples with noise and conditional vectors as input and trains adversarially against the discriminator. During the detection phase, if the discriminator determines that the input data is abnormal, the public opinion group to which the data belongs is marked as an abnormal sentiment group. The module sends the feature information of the abnormal group to the cross-context nested causal chain business opportunity discovery module and the business opportunity warning output module for corresponding risk warnings and in-depth analysis.

[0023] Preferably, the causal graph generative learning engine uses the difference between the predicted results of market indicators and the actual observed data from the causal relationship graph as a feedback signal, and adopts an unsupervised structured learning algorithm to dynamically adjust the causal graph; the causal graph generative learning engine further combines a generative adversarial training mechanism, regards the entire process from public opinion to market analysis as a generative model, evaluates the matching degree between the system output and the real situation by the discrimination process, and continuously optimizes the parameters of each functional module accordingly.

[0024] Preferably, the business opportunity early warning output module provides an interpretable causal chain basis and quantitative evaluation results for each potential business opportunity, including key public opinion events involved in the business opportunity, the market indicators affected, the expected time lag and magnitude of the impact, and the action plan suggested by the system for the business opportunity; the business opportunity early warning output module supports displaying business opportunities and early warning information to users through a visual interface, and allows users to provide feedback to confirm or adjust the analysis results, thereby incorporating human expertise into the causal graph generative learning engine to further optimize system performance.

[0025] A deep learning-based method for international market sentiment analysis and business opportunity mining, applied to the deep learning-based international market sentiment analysis and business opportunity mining system described above, specifically includes the following steps:

[0026] Step 1: Acquire public opinion text data from multiple countries and languages, and preprocess the data; collect EEG signals from some samples during the training phase for emotion model calibration;

[0027] Step 2: Input the pre-processed public opinion texts in each language into the pre-trained cross-cultural emotion recognition model, and output emotion representation and emotion intensity index with a unified emotion scale;

[0028] Step 3: Based on sentiment representation and relevant historical market data, construct a multi-granularity public opinion-market causal relationship graph; use graph neural networks and causal inference algorithms to update the node representations and causal connections of the causal graph, and identify the implicit causal paths in the evolution of public opinion;

[0029] Step 4: For the identified important public opinion events or links, conduct counterfactual scenario simulations to predict changes in market indicators under different assumptions; combine the causal link structure to explore potential cross-contextual supply or demand gap business opportunities and assess their impact.

[0030] Step 5: Monitor the emotion data stream and use generative adversarial networks to identify abnormal emotion groups; for detected anomalies, generate early warning information in a timely manner and include them in the scope of causal analysis;

[0031] Step 6: Output public opinion analysis report and business opportunity mining results for users' decision-making reference; at the same time, collect actual market post-hoc data and user feedback, feed them into the causal graph generative learning engine, update the causal graph neural network analysis module and abnormal emotion group detection module, and realize continuous adaptive learning.

[0032] Preferably, the training of the cross-cultural emotion recognition model in step 2 includes: using samples containing EEG signals, calculating the mean squared error loss between the emotion representation output by the text model and the emotion features extracted by EEG, and minimizing it by weighted merging with the conventional emotion classification loss, so as to obtain model parameters that can accurately predict emotions across languages. During inference, the input of public opinion text in any language is output as a vector representation in the common emotion space.

[0033] Step 3, constructing the causal relationship graph, includes: conducting correlation analysis on public opinion sentiment indicators and market indicators at different time scales to screen candidate causal relationships with time-series dependence; using a structured graph neural network to evaluate the candidate causal relationships, eliminating false associations and retaining true causal relationships to obtain a multi-level connected causal network; and continuously using new data to correct the causal network during system operation, so that the causal relationship graph can evolve dynamically.

[0034] Step 4, the counterfactual scenario simulation, includes: determining the current state parameters of the target public opinion event; setting at least one hypothetical scenario's future evolution parameters for the event; simulating the changes in public opinion and related market indicators step by step from the initial conditions of the hypothetical scenario based on the causal relationship graph; recording the predicted trajectories of key indicators under different scenarios and conducting comparative analysis to obtain scenario differences and key influencing factors for business opportunity assessment.

[0035] Step 5, abnormal sentiment group detection, includes: inputting real-time public opinion sentiment data into a pre-trained generative adversarial network discriminant model and outputting the probability value of its belonging to the normal mode; if the probability value is lower than a set threshold, the corresponding public opinion group is determined to be an abnormal sentiment group; recording the attributes and emotional characteristics of the abnormal group and outputting warning information; and further using the generative model to simulate and verify the abnormal mode in order to improve the confidence of the abnormal detection results.

[0036] Step 6, continuous adaptive learning, includes: comparing the results of the counterfactual sentiment simulation with the actual market posterior data, calculating the error and adjusting the weights or structure of the corresponding causal edges in the causal graph accordingly; adopting the results of anomaly detection, incorporating abnormal public opinion events into the nodes of the causal graph, and attempting to establish causal relationships between them and other nodes; and correcting the model output based on user feedback, adding or deleting corresponding relationships in the causal graph to continuously improve the accuracy of system analysis and prediction.

[0037] Beneficial Effects: The international market sentiment analysis and business opportunity mining system based on deep learning provided by this invention has the following beneficial effects:

[0038] 1. By collaborating with a cross-cultural emotion recognition module and a multi-source public opinion data acquisition module, this invention solves the technical problems of inconsistent emotion dimensions, semantic bias, and cross-domain incomparability in multilingual, multi-platform, and multicultural contexts. The cross-cultural emotion recognition module outputs emotion features within a unified representation space, ensuring that public opinion data from different language sources have consistent dimensions and stable distribution before entering downstream modeling. This achieves stability and comparability of emotion recognition under cross-language and cross-platform conditions, significantly reducing the impact of cross-domain bias on the accuracy and robustness of subsequent modeling.

[0039] 2. By linking the causal graph neural network analysis module, the emotion counterfactual simulation module, and the cross-context nested causal chain opportunity discovery module, this invention solves the technical problems of correlation mining's inability to provide causal explanations and its inability to clearly define intervention timing and business opportunity implementation paths. The causal graph neural network analysis module establishes multi-granularity causal relationships between the micro-event layer, the theme layer, and the macro-indicator layer; the emotion counterfactual simulation module outputs scenario trajectories and key intervention windows based on the causal structure; and the cross-context nested causal chain opportunity discovery module maps stable causal chains into actionable business opportunity clues and action suggestions. This enables causal chain-driven business opportunity positioning and timing-based decision-making. The technical effect is reflected in the ability to output actionable solutions with causal basis, including intervention windows and indicator increments, avoiding the shortcomings of traditional correlation methods that only provide relevant reminders but cannot guide action.

[0040] 3. By employing a closed-loop collaboration between the abnormal emotion group detection module and the causal graph generative learning engine, and triggering feedback at the business opportunity warning output module, this invention solves the technical problems of model degradation caused by rapid evolution of online contexts, high costs of manual re-annotation, and poor warning stability. The abnormal emotion group detection module identifies abnormal topic groups and regional populations in real-time streams. The causal graph generative learning engine jointly updates the emotion model, causal structure, and discriminator based on prediction and observation biases and feeds them back to the front-end inference, enabling the system to achieve adaptive convergence and structural correction under unsupervised or weakly supervised conditions. The technical effects are reflected in the significant suppression of false positives and false negatives during long-term online operation, rapid adaptation to emerging topics, and low maintenance and annotation costs, ensuring that the business opportunity warning output remains stable and reliable in dynamic contexts. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0042] Figure 1 This is a block diagram of the system structure of the present invention;

[0043] Figure 2 This is a flowchart of the cross-cultural emotion calibration process of the present invention;

[0044] Figure 3 This is the multi-granularity cause-effect graph of the present invention;

[0045] Figure 4 This is an emotional counterfactual simulation diagram of the present invention;

[0046] Figure 5 This is a detection chart for groups exhibiting abnormal emotions.

[0047] Figure 6 This is a schematic diagram of the closed loop of causal graph generative learning. Detailed Implementation

[0048] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0049] like Figure 1 As shown, a deep learning-based international market sentiment analysis and business opportunity mining system includes a multi-source sentiment data acquisition module, a cross-cultural sentiment recognition module, a causal graph neural network analysis module, a sentiment counterfactual simulation module, a cross-context nested causal chain business opportunity discovery module, an abnormal sentiment group detection module, a causal graph generative learning engine, and a business opportunity early warning output module. (See also...) Figure 1The overall system architecture and the data and feedback relationships between modules of this invention are shown in the figure. During system operation, the system first acquires the input data required for analysis through a multi-source data acquisition module, including but not limited to:

[0050] Global multilingual public opinion text data, covering social media posts, news articles, forum comments, e-commerce reviews, etc., is captured from multiple channels and categorized by language or region.

[0051] Supplemental EEG data was collected during the model training phase. Subjects from diverse cultural backgrounds were optionally recruited to read or view selected public opinion samples, and their emotional responses were recorded using EEG devices. This data was used to guide the training of the cross-cultural emotion model but was not required input for online operation.

[0052] The collected data, after cleaning and preprocessing, enters the cross-cultural emotion recognition module. This module performs sentiment analysis on each text, outputting a uniformly scaled emotion representation (e.g., emotion probability distribution or emotion vector). This step incorporates guidance from electroencephalogram (EEG) signals, ensuring the model's emotion recognition results across different languages ​​are comparable. The identified emotion time series is then input into the causal graph neural network analysis module.

[0053] The causal graph neural network analysis module maintains a causal graph where nodes represent events or indicators of different granularities (such as the sentiment value of a single article, the overall public opinion index for a given day, or the market sales of a particular industry), and directed edges represent potential causal influences. As new public opinion data arrives, the causal graph neural network analysis module uses the MG-CGNN model to update node representations and identify and verify causal relationships. For example, when changes in public opinion on a particular topic have a statistical dependency on subsequent market indicator fluctuations, the causal graph neural network analysis module uses a causal inference algorithm to determine whether a causal relationship exists and adds a causal edge to the graph. MG-CGNN allows connections between nodes of different scales, thus forming a hierarchical causal network (such as micro-events influencing macro-trends, and macro-background modulating the effects of micro-events).

[0054] Based on the causal graph construction, the emotion counterfactual simulation module is activated periodically or when a surge in public opinion is detected. It reads the subgraph of a specific topic from the current causal graph and generates emotion evolution trajectories under several hypothetical scenarios based on the causal model. For example, given the current growth rate of public opinion, the simulator can assume two scenarios: "if the current trend continues" and "if the popularity suddenly doubles," respectively, and calculate the future change curves of the sentiment index. Furthermore, using the paths connecting this public opinion to market indicators in the causal graph, the simulator predicts the future values ​​of relevant market indicators (such as stock prices and sales volume) under each scenario, thus generating market impact results corresponding to different scenarios. The emotion counterfactual simulation module provides these counterfactual scenarios and their results to the cross-context nested causal chain opportunity discovery module for comparative analysis.

[0055] Meanwhile, the cross-context nested causal chain opportunity discovery module continuously monitors the outputs of the causal graph neural network analysis module and the sentiment counterfactual simulation module to uncover potential business opportunities. The cross-context nested causal chain opportunity discovery module performs two tasks: First, based on the current structure of the causal graph, it searches for causal chains spanning different contexts. For example, if the causal graph contains a path from "high social media sentiment (US)" through "product word-of-mouth spread (Europe)" to "supply shortages leading to price increases (global)," the module identifies it as a cross-context nested causal chain and extracts the implied business opportunity—that is, the relevant product may be in short supply in Europe. Second, it compares the results of multiple scenarios from the sentiment counterfactual simulation module, identifying the indicators with the greatest differences (such as a surge in sales in one scenario while remaining stable in another). Based on this, it judges the potential impact of a decision or event (such as whether to launch a new product or take intervention measures) on the market, thereby identifying business opportunities or risks. This counterfactual analysis enables the cross-context nested causal chain opportunity discovery module to not only focus on causal chains that have already occurred, but also to uncover potential opportunities that "have not occurred but are possible".

[0056] In the abnormal emotion group detection module, the system simultaneously outputs emotion data from the cross-cultural emotion recognition module and feeds it to the abnormal emotion group detection module for abnormal pattern analysis. The abnormal emotion group detection module contains a semi-supervised GAN model: the discriminator learns to distinguish between normal and abnormal emotion patterns, while the generator attempts to produce convincingly realistic normal emotion samples. During actual operation, the abnormal emotion group detection module feeds real-time public opinion emotion data to the discriminator. If the data distribution of a certain group significantly deviates from the normal distribution that the generator can produce, the discriminator will classify it as abnormal. For example, if most regions are neutral on a certain event, but a small group exhibits extreme anger, then that group is considered an abnormal emotion group. The abnormal emotion group detection module outputs the characteristics of these abnormal groups (involving topics, regions, and demographic attributes), notifying the cross-context nested causal chain business opportunity discovery module—because abnormal emotions often indicate opportunities or hidden dangers that have not been noticed by the public—and directly sending early warning information to the business opportunity warning and result output module to attract the attention of decision-makers. In addition, the generator of the abnormal emotion group detection module can also generate similar emotion data based on the identified abnormal patterns to expand the training set, so that the discriminator can continuously improve its sensitivity to detecting abnormal patterns.

[0057] The entire system is coordinated and driven by a causal graph generative learning engine. This engine is not an independent process step, but rather a strategy layer that runs through all modules. It periodically (or after significant events) evaluates overall system performance metrics (such as the accuracy of emotion recognition, the validation results of causal predictions, and the difference between simulated predictions and actual outcomes), and updates the model parameters of each module through adaptive algorithms. For example, the causal graph generative learning engine compares the predictions of the emotion counterfactual simulation module with subsequent real market data. If it finds deviations in certain causal chain inferences, it uses these errors to adjust the weights or confidence levels of corresponding edges in the multi-granularity causal graph neural network analysis module. Furthermore, the engine incorporates new abnormal patterns identified by the abnormal emotion group detection module into causal learning, determining whether they represent new causal factors and thus deciding whether to add new nodes or edges to the causal graph. The engine also manages the continuous learning of the cross-cultural emotion recognition module, updating the emotion model to adapt to the latest changes in the linguistic environment when new EEG training data or credible user feedback is collected. Through this iterative training under unsupervised and weakly supervised feedback, the causal graph generative learning engine achieves "learning while running," continuously improving the system's inference accuracy regarding complex public opinion-market relationships.

[0058] Finally, all analysis results are aggregated in the Business Opportunity Early Warning and Results Output module. This module integrates and presents to the user a list of potential opportunities from the cross-contextual nested causal chain business opportunity discovery module, warning signals from the abnormal sentiment group detection module, and other relevant information. Output formats may include: charts (e.g., comparison curves of trending public opinion and related market indicators), causal chain diagrams, text descriptions (e.g., "Demand for a certain product is expected to surge by 20% in market X next quarter"), and alerts (e.g., "Warning: Young people are showing abnormally negative sentiment towards brand Y"). Users can then further inquire about details or obtain decision-making advice. It is worth noting that each business opportunity output by this system is accompanied by explainable evidence, such as the corresponding causal chain and key influencing factors, enabling decision-makers to understand the source of the opportunity and verify it. This explainability is attributed to the system's core causal analysis, which differs from traditional black-box prediction algorithms.

[0059] The working principles of each module are as follows:

[0060] The cross-cultural emotion recognition module is one of the fundamental modules of this invention, designed to address the issue of inconsistent emotion measurement across different languages. This module employs a deep learning model (e.g., a Transformer-based sentiment analysis network) and performs cross-cultural calibration by fusing electroencephalogram (EEG) data, ensuring comparability of emotion results for text outputs in different languages.

[0061] See Figure 2 The training and inference pathways for cross-cultural emotion calibration are shown in the figure. The model of the cross-cultural emotion recognition module receives text input (which can be sentences, paragraphs, etc.) and outputs emotion representations. Emotion representations can be multi-dimensional continuous vectors (such as emotion embeddings) or discrete distributions (such as positive, neutral, and negative probabilities). The model structure consists of two parts: a multilingual text encoding layer and an emotion prediction layer. The text encoding layer uses a multilingual pre-trained model (such as multilingual BERT) to map texts from different languages ​​to a unified semantic space; the emotion prediction layer converts the semantic representation into an emotion representation. Unlike ordinary sentiment analysis models, this module adds an EEG feature alignment submodule: during the training phase, a small number of samples containing EEG annotations are introduced, and an alignment loss function is used to align the model's text emotion output with the emotion features extracted from the EEG signals.

[0062] In the specific training process, for training samples with EEG records, assuming the text of the i-th sample is... The corresponding EEG signal time series is First, analyze the brainwave signals. Through feature extraction network Extract feature vectors reflecting emotional responses (For example, using emotion features formed by wavelet energy, frequency band power, etc.). The text portion of the cross-cultural emotion recognition module outputs an emotion vector. One of the training objectives is to minimize and The difference between them can be defined using the mean squared error loss as follows:

[0063]

[0064] Where N is the number of training samples with EEG annotations. It is a model for text The emotional characteristics of the output This refers to the emotional features mapped from the corresponding electroencephalogram (EEG) signals. By minimizing this loss, the model can learn to make the text's emotional representation as close as possible to the actual physiological emotional response. It should be noted that in the above formula... It represents the true value of emotion under EEG, and its acquisition method can be defined according to specific experiments, such as mapping EEG signals to scores of emotional dimensions (excitement, pleasure).

[0065] In addition, the cross-cultural emotion recognition module is trained using conventional emotion-annotated data (such as manually annotated emotion categories) and employs cross-entropy loss. Improve the model's basic sentiment classification ability. The total training loss can be summarized as follows: ,in This serves as a weighting factor. Through this training strategy, the model not only correctly classifies sentiment polarity but, more importantly, resolves the issue of inconsistent sentiment scales across languages. For example, for a Chinese sentence and an English sentence expressing similar levels of anger, a standard model might assign different sentiment scores due to language differences. However, by introducing EEG calibration, this module can output more consistent anger intensity values.

[0066] After deployment, the cross-cultural sentiment recognition module performs the same processing flow on newly input sentiment texts in various languages. Since the sentiment spaces of different languages ​​were aligned during training, the output results can be directly compared. For example, when comparing the sentiment intensity of two sentiments, "American users complain about delayed product delivery" and "Japanese users complain about delayed product delivery," the anger scores output by the cross-cultural sentiment recognition module are comparable, ensuring that subsequent causal analysis will not be biased due to differences in cultural expression. This lays a solid foundation for the reliability of sentiment data for the entire system.

[0067] It is worth mentioning that the EEG signals used in this module only need to be collected in small quantities during the training phase. In practical applications, ordinary public opinion texts do not require any additional physiological signals, which ensures the feasibility and wide applicability of the system. Through the cross-cultural emotion recognition module, this invention establishes a unified emotion measurement standard globally, providing reliable input data for subsequent causal analysis and business opportunity mining.

[0068] The multi-granularity causal graph neural network module is the core analysis module of this invention. It is responsible for mining the deep causal relationships in the evolution of public opinion and representing them in the form of a causal graph. In its implementation, this module adopts a technical approach that combines graph neural networks (GNNs) with causal reasoning, enabling it to capture the causal relationships between events at different granular levels.

[0069] The multi-granularity causal graph neural network module maintains a dynamic graph structure. The causal graph is structured as follows: The node set V represents related entities or events, with different nodes belonging to different granularity levels. The edge set E represents causal relationship assumptions. For example, nodes could include "daily sentiment index of a certain brand of mobile phone on Twitter," "daily closing price of the brand's stock," or "macroeconomic index of a certain region." The granularity ranges from daily sentiment on specific platforms to macroeconomic indicators. The initial causal graph can be initialized using background knowledge or expert experience, such as connecting candidate causal edges from "rising sentiment index" to "rising stock price" based on common sense. As the system operates, the causal graph structure will dynamically adjust and expand.

[0070] The MG-CGNN model operates on causal graphs to uncover latent causal relationships. It consists of multi-layer graph neural network units and causality discriminant units. The graph neural network units iteratively update the embedding representation of each node based on the current graph structure. The first unit integrates information from neighboring nodes to learn the representation of related relationships; the second unit uses causal inference methods (such as inference based on comparative experiments or Do-Calculus) to evaluate whether an edge truly has a causal impact, rather than merely being correlated. Specifically, for each pair of directed connections... The model attempts to assess, after controlling for confounding factors, whether a change in u significantly leads to a change in v. If so, it assigns an edge... A higher causal weight is applied; otherwise, its weight is reduced or the edge is removed.

[0071] MG-CGNN is called "multi-granular" because it introduces a mechanism of hierarchical grouping and multi-scale time windows:

[0072] like Figure 3As shown, in terms of hierarchical structure, the MG-CGNN model divides nodes into different groups according to granularity (e.g., micro-event groups, macro-indicator groups). When the graph neural network computes message passing, it allows information flow across groups and also supports intra-group aggregation to extract patterns at different levels of abstraction. This allows it to capture nested causal relationships such as how details affect the whole and how the overall trend constrains the details.

[0073] In terms of time scale, the MG-CGNN model considers different time delays and window lengths when constructing causal relationships. For example, for each pair of nodes... The model will try different time delays. Come to check At any moment Changes exist The existence of such effects can be explored at the hourly, daily, weekly, and monthly levels. This approach can uncover both short-term, immediate causal effects and long-term, lagged effects. For example, different scales of causality, such as "sudden negative public opinion → stock price drop on the same day" and "decline in consumer confidence → decline in quarterly sales," can be detected.

[0074] The training of MG-CGNN is primarily conducted adaptively during system operation, without requiring a fully supervised dataset, which is one of the innovations of this invention. In the initial stage, historical data can be used to calibrate training on known causal relationships. For example, causal pairs tested using historical statistics can guide model parameters. Subsequently, the model increasingly employs unsupervised structural learning algorithms: iteratively optimizing the graph structure and node embeddings by minimizing prediction errors (e.g., using a causal graph to predict market indicators and calculating errors based on actual values) and introducing sparsity regularization (preferring fewer causal edges). Whenever new data arrives, the model re-evaluates the confidence of each candidate causal relationship based on the new data, adjusting the causal graph accordingly. For example, if the event node "approaching holidays" is recently strongly correlated with "search popularity of a certain product" and follows a causal direction (the former changes first, the latter changes later), the model adds a causal edge to represent the causal relationship where holiday factors lead to increased demand. Conversely, if a hypothetical causal edge in the original graph has not received data support for a long time, the model weakens its weight or removes it, thus maintaining the accuracy of the causal graph.

[0075] As MG-CGNN continues to operate, we can obtain an ever-improving causal knowledge graph of public opinion. This graph intuitively reveals the causal structure hidden behind massive amounts of data. For example, it can tell us a clear chain such as "increased positive social media reviews → improved brand reputation → increased sales," as well as the strength and conditions of each step. Simultaneously, for a given target node (such as a specific market indicator), MG-CGNN can trace its possible multi-source factor combinations, that is, identify the various public opinion factors that may lead to changes in that indicator. These outputs will be directly used by the cross-contextual nested causal chain opportunity discovery module to locate the causal path of business opportunities. The results of MG-CGNN can also be provided to analysts for review as an interpretable AI tool. For example, in the face of a sales slump, it can point out the underlying causal chain (perhaps "negative review videos → decreased consumer trust → sales decline"), helping companies take timely measures at specific links in the chain.

[0076] In summary, the multi-granularity causal graph neural network module enables the automatic mining of causal relationships between public opinion and market data, which largely overcomes the limitations of manual analysis and provides solid technical support for this invention to uncover hidden business opportunities.

[0077] The Emotion Counterfactual Simulation module focuses on predicting the future and analyzing hypothetical scenarios. It utilizes a causal graph model generated by a multi-granularity causal graph neural network analysis module to simulate the development of current public opinion along multiple paths, thereby assessing potential market trends under different scenarios. This module essentially provides decision-makers with a testing ground, allowing them to preview outcomes before real events occur and providing a reference for strategy formulation.

[0078] The simulations in the emotion counterfactual simulation module do not occur without cause, but rather have explicit triggering conditions. For example:

[0079] When the intensity of public sentiment or the scope of a public opinion event is detected to be rising abnormally (exceeding a set threshold), it is considered that the event has the potential to become a hot topic, and a simulation is triggered.

[0080] Regular forecasting of demand, such as simulating market sentiment for the next day after the market closes each day, or forecasting hot topics for the following week each week.

[0081] Users can manually specify scenarios of interest, such as requesting a simulation of "what impact would a sudden policy release have on the market."

[0082] The emotion counterfactual simulation module uses a cause-and-effect graph as its basis. The specific steps are as follows:

[0083] a) First, extract a subgraph related to the target sentiment or event from the overall causal graph. This includes the sentiment node, its directly and indirectly affected nodes (e.g., related market indicator nodes), and the causal paths connecting these nodes. This subgraph reflects how the sentiment we are interested in influences the market structure through causal chains.

[0084] b) Identify the different scenarios to be simulated. This typically includes at least one baseline scenario (e.g., the current trend continuing) and one or more hypothetical scenarios (e.g., an event X suddenly occurs at time t in the future, or the growth rate doubles). Scenario parameters can be quantitatively represented, for example, "the sentiment index will maintain a daily growth rate of +5% for the next 3 days" vs. "the sentiment index will suddenly increase by 20% tomorrow and then fall back."

[0085] c) Based on the causal subgraph, perform time-step simulations for each scenario. For each time step, under the assumed scenario conditions, calculate the sentiment value at the sentiment point, and then proceed along the causal graph, calculating the changes in affected downstream nodes (market indicators) based on causal relationships. For example, if the causal graph has an edge representing "sentiment index → ​​stock price", and the causal effect model... Given this, then based on the simulated emotional changes... Calculate stock price changes This process is repeated step by step, calculating the entire sub-causal graph. The quantified effects in the causal graph (which may be estimated from historical data) are used here.

[0086] d) Repeat the above process to obtain the future time series trajectories of key indicators (such as sentiment index and market indicators) for each scenario. For example, the public opinion sentiment curve and the corresponding stock price curve for scenario A, and the similar curve for scenario B, etc.

[0087] After obtaining simulation results from multiple scenarios, the emotion counterfactual simulation module compares these results and extracts information that is meaningful for business opportunities:

[0088] Identify the indicators with the greatest difference between scenarios. For example, if market sales are stable in the baseline scenario but increase significantly in the hypothetical scenario, it indicates that if the hypothetical scenario occurs, it will bring sales opportunities.

[0089] Identify common risks across scenarios. If a brand's reputation continues to decline regardless of changing scenarios, it indicates that intervention alone is insufficient to avoid the risk, and a more fundamental strategy is needed.

[0090] Key turning point analysis involves identifying significant divergences in indicator patterns before and after a specific point in time under different scenarios during the simulation. This suggests that this point may be a crucial window for intervention decisions. For example, if the simulation shows that releasing a new product next Monday can reverse public opinion, otherwise it will spiral out of control, then next Monday is a critical point.

[0091] like Figure 4 As shown, in one embodiment, the emotion counterfactual simulation module constructs two types of counterfactual samples based on the current public opinion state: Scenario A (continuous escalation of public opinion) and Scenario B (timely intervention to cool down public opinion), and generates corresponding trajectories A and B. The system compares the results of the two trajectories at the indicator level and outputs Δsales volume, Δprice, and key intervention points and windows of opportunity to assist in decision-making.

[0092] The emotion counterfactual simulation module sends the information obtained from the above analysis to the cross-context nested causal chain opportunity discovery module. Based on this, the cross-context nested causal chain opportunity discovery module can update its judgment of the current opportunity, such as confirming that "if action X is taken, there is opportunity Y; otherwise, there will be a loss Z." The purpose of counterfactual simulation is to help the system see not only established facts but also potential possibilities. For businesses, this means gaining a competitive advantage: preparing for a possible trend before competitors even realize it.

[0093] In terms of implementation, the emotion counterfactual simulation module can employ a variety of technical means:

[0094] Generative models (such as Variational Autoencoders (VAEs) or Temporal GANs) are used to generate future sequences. These models are conditioned on causal graphs and treat future sentiment and indicator sequences as the data to be generated.

[0095] Based on equation modeling, the results are calculated directly using structural equations formed by causal relationships (which requires quantitative priors of causal effects or estimation from data).

[0096] Alternatively, Monte Carlo simulation can be used to superimpose random perturbations to obtain the possible distribution range.

[0097] Regardless of the method, the key is to use causal graphs to ensure that the simulation follows causal constraints, rather than arbitrary statistical predictions. Only in this way can the simulation results have credibility and interpretability. For example, traditional time series forecasting may predict stock prices, but it cannot explain why they rise; while causal-driven simulations can point out the logic that "improved consumer sentiment led to increased purchases, which drove up stock prices."

[0098] The cross-context nested causal chain opportunity discovery module integrates information from the aforementioned modules, acting as an analyst to extract truly commercially meaningful insights and actionable recommendations. In particular, it emphasizes the use of cross-context nested causal chains, aiming to discover opportunities across different language and cultural backgrounds and industry sectors.

[0099] The cross-context nested causal chain business opportunity discovery module first extracts complete causal chains from the causal graph. A causal chain is a series of interconnected causal relationships, such as... This means A leads to B, B leads to C, and C leads to D. Due to the large number of nodes in the causal graph, there may be numerous paths. The cross-context nested causal chain opportunity discovery module extracts information for specific targets: the target could be a market indicator (such as product sales) or an anomalous event node provided by the abnormal sentiment group detection module. The algorithm uses depth-first or breadth-first search, starting from the target node and tracing back along the causal edges to find its cause, continuing upwards until a source without antecedents is found (the source could be external events such as policies or holidays, or factors that the model can no longer explain). This upward tracing forms a causal tree, and then suitable paths are selected as candidate causal chains. For example, if the target is "product X sales decline," backward tracing might find the path: "Sales decline ← surge in negative consumer reviews ← product defect exposure ← opinion leader reports." This is a cross-domain chain, from media reports (public opinion) to product reputation and then to sales results.

[0100] In the extracted causal chains, the cross-context nested causal chain opportunity discovery module pays particular attention to cross-contextual (cross-language, cross-regional, cross-industry) features. That is, if nodes in a chain belong to different contexts, the chain is marked as a cross-context nested causal chain. For example, in the above example, if "opinion leader reporting" occurred in US media, while "a surge in negative consumer reviews" occurred in user comments around the world, then the chain crosses geographical contexts. Similarly, a chain may cross industries (such as social media events affecting raw material prices). The cross-context nested causal chain opportunity discovery module detects such cross-context jumps through the node's metadata (language tags, regional tags, topic types, etc.). After identifying a cross-context chain, the system considers it a significant clue to a global business opportunity or risk. For example, a chain like "a country's policy change → rising international crude oil prices → global shortage of related plastic products" suggests a global supply gap opportunity.

[0101] To transform causal chains into concrete business opportunities, the cross-context nested causal chain business opportunity discovery module also introduces a series of evaluation rules:

[0102] By comprehensively considering the confidence levels of each side of the causal chain (provided by the causal graph) and the frequency of similar historical occurrences, a confidence score is assigned to the chain. Only chains with high confidence levels are used for business opportunity inference.

[0103] Estimate the final impact on the target by quantifying the causal chain effect. For example, calculate "a change of X% in the source event ultimately leads to a change of Y% in the target indicator". If Y is large enough, it signifies a significant business opportunity.

[0104] Estimate the time required for the chain of events to take effect and determine if it falls within a window of opportunity. If an opportunity is found to be rapidly disappearing (missed opportunity) or taking a long time to materialize (not urgent), its priority is reduced.

[0105] Determine if there are any links in the chain that a company can intervene in. If a company can take action to influence a link, then the business opportunity can be seized. For example, if the chain involves a competitor's mistake, the company can take advantage of it; if it is caused by an uncontrollable natural disaster, it can only serve as a risk warning.

[0106] Using the above rules, the cross-contextual nested causal chain opportunity discovery module filters and sorts causal chains, ultimately selecting the most commercially valuable chains as business opportunities. For each business opportunity chain, the system generates a corresponding description, such as, "Due to event A occurring in country X, there is a global supply shortage of product B. It is recommended to increase inventory in country X to meet potential demand increases." This essentially translates the chain into human-understandable and actionable strategic suggestions.

[0107] The cross-context nested causal chain business opportunity discovery module will also incorporate information from the emotion counterfactual simulation module and the abnormal emotion group detection module into the business opportunity judgment:

[0108] If simulations show a greater opportunity in a particular scenario, the cross-context nested causal chain opportunity discovery module may suggest actions to facilitate that scenario. For example, if the simulation indicates that "investing in marketing will significantly improve public sentiment and boost sales," it might suggest "actively marketing to guide public opinion." Anomaly detection results would indicate "niche opportunities" or "potential crises." The cross-context nested causal chain opportunity discovery module generates separate analysis reports for groups exhibiting abnormal sentiment. For example, if it detects that a niche market group has an extreme preference for a certain accessory, this represents a niche market opportunity that has not yet been noticed by large companies, and early investment is advisable.

[0109] By integrating information from multiple sources, the business opportunity discovery module can output diverse and in-depth insights, including a list of market opportunities, a list of risk warnings, and action suggestions, which are then presented to users by the business opportunity alert and results output module. Behind these insights are causal chains identified by the system, enabling users to understand why these opportunities exist when considering the recommendations.

[0110] The causal graph generative learning engine is the core of intelligent learning throughout the entire system. It is not a standalone function, but a framework that integrates system operation and model learning. Its task is to enable the system to continuously learn and optimize from new data under unsupervised or weakly supervised conditions, thereby adapting to changing public opinion and market environments.

[0111] like Figure 6 As shown, the causal graph generative learning engine adopts a generative learning approach, using the causal graph model as a knowledge container. It obtains feedback signals through generation simulations and comparative inferences, thereby adjusting the model. The entire process resembles a loop: initial causal model → generating predictions or simulations based on the model → comparing with real-world data → revising the model → using the new model for generation again →... continuously closing the loop.

[0112] For example, in counterfactual emotion simulation, a causal graph generative learning engine compares the simulated results of a public opinion event with the actual subsequent development. If the simulation deviates significantly, it indicates that the causal graph incorrectly grasps the causal mechanism of the event. The causal graph generative learning engine then adds or modifies the causal relationship assumptions accordingly: perhaps a key influencing factor is missing, or the influence of a certain path is overestimated. Such adjustments are achieved through algorithmic means, such as using Bayesian inference-based methods to correct the causal network structure.

[0113] In practical applications, most events lack explicit labels that tell the system which event caused which. Causal graph generative learning engines are driven by observing unsupervised signals such as changes in data distribution and model prediction errors. For example, if a product's sales decline continuously while negative comments on social media surge, although the correlation isn't explicitly stated, the model notices a common statistical pattern and attempts to establish a connection in the causal graph. If, after establishing the connection, it is validated over time that this better explains the data (e.g., reduces prediction error), then the connection is retained.

[0114] While primarily unsupervised, causal graph generative learning engines can also incorporate some human feedback or external knowledge. For example, if a user confirms or denies a causal relationship, the system uses this as a strong constraint to adjust the model. Similarly, if new trends mentioned in industry reports are imported as prior knowledge, the model adds corresponding nodes and connections. The design of causal graph generative learning engines allows for the integration of knowledge from multiple sources while ensuring model consistency.

[0115] As shown in the figure, the causal graph generative learning engine borrows the idea of ​​GAN, transforming part of the learning task into a generative adversarial process:

[0116] On the one hand, the system generator (which can be understood as a positive analysis process that includes a cross-cultural emotion recognition module, a multi-granularity causal graph neural network analysis module, an emotion counterfactual simulation module, and a cross-context nested causal chain business opportunity discovery module) continuously generates analysis and prediction results for the data;

[0117] On the other hand, an implicit discrimination process continuously checks whether the system output matches the real world (e.g., discriminating the difference between the predicted market trend and the actual market data distribution).

[0118] The errors in the discrimination process, in turn, guide the generator updates. For example, if the system consistently fails to predict a certain type of market fluctuation, and the discriminator issues a high false signal for this type of output, then the generator needs to adjust its internal causal model to correct the prediction. This idea permeates causal graph generative learning engines, enabling them to move beyond passive learning and actively engage in trial and error correction.

[0119] To prevent model drift or oscillation caused by continuous learning, the causal graph generative learning engine sets monitoring metrics, such as whether prediction performance has steadily improved over a recent period and whether the frequency of changes in causal relationships has decreased. When model training converges, the system enters a stable operating state, making adjustments only for significant new information. This ensures the consistent reliability of the system output and prevents the introduction of noise due to overlearning.

[0120] By employing a unified causal graph generative learning framework, the system of this invention possesses self-evolution capabilities. Over time, it becomes increasingly adept at understanding the complex relationships between public opinion and the market across different cultural contexts. For example, initially, the system may not understand the public opinion impact of a particular emerging social media platform, but after a period of data learning, the causal graph generative learning engine will capture its influence and incorporate it into the causal model. Ultimately, this continuous learning enables the system to adapt effectively to changes and is difficult to imitate, thus creating a technological barrier.

[0121] The abnormal sentiment group detection module aims to solve the problem of early detection of weak signals. By introducing Generative Adversarial Network (GAN) technology, it greatly improves the sensitivity and accuracy of identifying abnormal public opinion patterns. It can be seen as the system's sentinel, monitoring those unusual corners in massive public opinion data.

[0122] like Figure 5 As shown, the abnormal emotion group detection module consists of a discriminator D and a generator G. The discriminator D takes an emotion data sequence as input and outputs the probability of judging whether it is normal or abnormal; the generator G takes noise and conditional information as input and outputs a simulated emotion data sequence, the purpose of which is to deceive the discriminator into thinking it is real normal data. Here, normal refers to conforming to the mainstream public opinion pattern, and abnormal refers to a rare pattern that deviates from the norm.

[0123] During the training phase, we can train the discriminator D using manually selected normal samples from historical data, while simultaneously training the generator G to produce similar normal samples. The training objective is for the discriminator D to be unable to distinguish between the data generated by the generator and the real normal data, i.e. Approximately true. Alternatively, semi-supervised training can be used, allowing the discriminator D to not only judge true / false, but also classify specific anomaly types (if some anomaly categories are known).

[0124] After training, discriminator D acquires anomaly detection capabilities. During online system operation, it aggregates public opinion sentiment data for each time window. (This could be a set of user sentiment samples or a sentiment sequence for a single topic), the discriminator calculates... If the probability is higher than the threshold, then it is considered... The corresponding group or topic exhibits an abnormal emotional pattern. The abnormal emotion group detection module will output a description of the anomaly, including: the characteristics of the abnormal group (region, population, topic, etc.), the type of abnormal emotion (such as extreme optimism, extreme anger), the time of occurrence, and the duration.

[0125] For example, in the reviews of an e-commerce platform, 95% of product reviews show a stable, neutral sentiment. However, a sudden surge of enthusiastic positive reviews appears from only one region speaking a less common language, which is unusual for the overall market. Discriminator D will detect this significant difference in sentiment distribution between the reviews from this less common language region and other regions, thus identifying it as an anomaly. The output might be: "An anomalous sentiment group detected: located in the South American market, Spanish-speaking region, showing unusually high positive sentiment towards product X." This suggests a potential business opportunity—the South American market has huge potential demand for product X and deserves attention.

[0126] The abnormal emotion group detection module not only detects anomalies but also uses a generator G to simulate abnormal patterns to aid in verification and subsequent processing. Specifically, when the discriminator D detects an abnormal pattern... We can input its features into the generator G to try to generate similar samples. .if It can successfully deceive the discriminator, that is If it's judged as normal, it means the abnormal pattern isn't prominent in the model's view and might be a false alarm; conversely, if... If it is still judged as anomalous, it indicates that G also struggles to simulate this pattern, and the anomaly is very obvious and realistic. Through G's experimentation, we can further evaluate the confidence level of the anomalous signal. Furthermore, the generator G can generate various variant data, which can be added to the training set to enhance the discriminator, thereby improving the model's sensitivity to detecting similar patterns (preventing similar patterns from being treated as new anomalies the next time they appear).

[0127] Once the abnormal emotion group detection module outputs an abnormal alarm, the system will automatically trigger the relevant processing:

[0128] The notification's cross-context nested causal chain opportunity discovery module checks whether there are chains in the causal graph involving the anomaly, and considers them as potential opportunities / risks. For example, if the anomaly group might be an emerging demand, the causal graph will follow up by expanding it as a new node, connecting it with sales figures, etc.

[0129] The business opportunity alert and results output module generates timely alert reports. As in the previous example, the report might state: "South American Hispanic-speaking users have shown unusual enthusiasm for X, which may indicate strong unmet demand in the region. We recommend paying attention to this."

[0130] The long-term accumulation of this anomalous data can also feed back into the cross-cultural emotion recognition module and the multi-granularity causal graph neural network analysis module. For example, if a unique pattern frequently appears in a certain culture, it can prompt the emotion model to adjust its baseline for that culture; frequently occurring anomalies will eventually be incorporated into the model as the new norm.

[0131] By enhancing anomaly detection with GANs, the system effectively solves the problems of high false negative and false positive rates associated with traditional rule-based and statistical methods. The introduction of GANs makes detection more flexible (the model learns normal distributions itself, rather than manually setting thresholds) and more intelligent (it can simulate data to verify hypotheses). Especially in cross-cultural scenarios, anomalous distributions may only be abnormal within a specific culture and not noticeable globally. Without a model to model each subgroup, these anomalies are easily overlooked. This invention successfully improves this capability by employing an abnormal emotion group detection module, ensuring the early discovery of potential disruptive opportunities.

[0132] In summary, through the close collaboration of the various modules, the international market sentiment analysis and business opportunity mining system of this invention can automatically learn and extract valuable business intelligence from massive amounts of cross-language data, helping enterprises to seize opportunities and mitigate risks in advance. In terms of implementation, each module is based on a mature deep learning framework with creative improvements, making it feasible. In terms of effectiveness, through the integration of these creative modules, the system significantly outperforms traditional sentiment analysis tools. This system can be deployed on enterprise intelligent decision support platforms, investment risk control platforms, etc., providing a scientific basis for relevant decisions.

[0133] A deep learning-based method for international market sentiment analysis and business opportunity mining, applied to the deep learning-based international market sentiment analysis and business opportunity mining system described above, specifically includes the following steps:

[0134] Step 1: Acquire public opinion text data from multiple countries and languages, and preprocess the data; collect EEG signals from some samples during the training phase for emotion model calibration;

[0135] Step 2: Input the pre-processed public opinion texts in each language into the pre-trained cross-cultural emotion recognition model, and output emotion representation and emotion intensity index with a unified emotion scale;

[0136] Step 3: Based on sentiment representation and relevant historical market data, construct a multi-granularity public opinion-market causal relationship graph; use graph neural networks and causal inference algorithms to update the node representations and causal connections of the causal graph, and identify the implicit causal paths in the evolution of public opinion;

[0137] Step 4: For the identified important public opinion events or links, conduct counterfactual scenario simulations to predict changes in market indicators under different assumptions; combine the causal link structure to explore potential cross-contextual supply or demand gap business opportunities and assess their impact.

[0138] Step 5: Monitor the emotion data stream and use generative adversarial networks to identify abnormal emotion groups; for detected anomalies, generate early warning information in a timely manner and include them in the scope of causal analysis;

[0139] Step 6: Output public opinion analysis report and business opportunity mining results for users' decision-making reference; at the same time, collect actual market post-hoc data and user feedback, feed them into the causal graph generative learning engine, update the causal graph neural network analysis module and abnormal emotion group detection module, and realize continuous adaptive learning.

[0140] Preferably, the training of the cross-cultural emotion recognition model in step 2 includes: using samples containing EEG signals, calculating the mean squared error loss between the emotion representation output by the text model and the emotion features extracted by EEG, and minimizing it by weighted merging with the conventional emotion classification loss, so as to obtain model parameters that can accurately predict emotions across languages. During inference, the input of public opinion text in any language is output as a vector representation in the common emotion space.

[0141] Step 3, constructing the causal relationship graph, includes: conducting correlation analysis on public opinion sentiment indicators and market indicators at different time scales to screen candidate causal relationships with time-series dependence; using a structured graph neural network to evaluate the candidate causal relationships, eliminating false associations and retaining true causal relationships to obtain a multi-level connected causal network; and continuously using new data to correct the causal network during system operation, so that the causal relationship graph can evolve dynamically.

[0142] Step 4, the counterfactual scenario simulation, includes: determining the current state parameters of the target public opinion event; setting at least one hypothetical scenario's future evolution parameters for the event; simulating the changes in public opinion and related market indicators step by step from the initial conditions of the hypothetical scenario based on the causal relationship graph; recording the predicted trajectories of key indicators under different scenarios and conducting comparative analysis to obtain scenario differences and key influencing factors for business opportunity assessment.

[0143] Step 5, abnormal sentiment group detection, includes: inputting real-time public opinion sentiment data into a pre-trained generative adversarial network discriminant model and outputting the probability value of its belonging to the normal mode; if the probability value is lower than a set threshold, the corresponding public opinion group is determined to be an abnormal sentiment group; recording the attributes and emotional characteristics of the abnormal group and outputting warning information; and further using the generative model to simulate and verify the abnormal mode in order to improve the confidence of the abnormal detection results.

[0144] Step 6, continuous adaptive learning, includes: comparing the results of the counterfactual sentiment simulation with the actual market posterior data, calculating the error and adjusting the weights or structure of the corresponding causal edges in the causal graph accordingly; adopting the results of anomaly detection, incorporating abnormal public opinion events into the nodes of the causal graph, and attempting to establish causal relationships between them and other nodes; and correcting the model output based on user feedback, adding or deleting corresponding relationships in the causal graph to continuously improve the accuracy of system analysis and prediction.

[0145] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based international market public opinion analysis and business opportunity mining system, characterized in that, The system comprises a multi-source public opinion data collection module, a cross-cultural emotion recognition module, a causal graph neural network analysis module, an emotion counterfactual simulation module, a cross-context nested causal chain opportunity discovery module, an abnormal emotion group detection module, a causal graph generative learning engine, and an opportunity early warning output module. The multi-source public opinion data collection module is used to obtain public opinion text data from different countries and languages, and obtain electroencephalogram data for cross-cultural calibration in the training stage. The cross-cultural emotion recognition module performs sentiment analysis on public opinion texts in different languages based on a deep learning model, and outputs emotion representations that can be compared across languages. The causal graph neural network analysis module uses a multi-granularity causal graph neural network to construct a causal relationship graph of public opinion and market events, and identifies implicit causal paths in the evolution of public opinion at different time and semantic scales. The emotion counterfactual simulation module simulates the emotion evolution and market impact of a target public opinion event under multiple scenarios using the causal relationship graph, compares the differences in market-related indicators under different hypothetical scenarios, and is used to predict potential market changes. The cross-context nested causal chain opportunity discovery module extracts cross-language and cross-regional nested causal chains from the causal relationship graph and emotion simulation results, analyzes global supply or demand changes triggered by public opinion, identifies potential business opportunities, and evaluates their impact and feasibility. The abnormal emotion group detection module monitors abnormal emotion pattern groups in public opinion data based on a generative adversarial network model, generates an early warning signal when an emotion group deviating from the normal pattern is detected, and provides abnormal emotion pattern features to the cross-context nested causal chain opportunity discovery module for further analysis. The causal graph generative learning engine receives information from the causal graph neural network analysis module, the emotion counterfactual simulation module, the abnormal emotion group detection module, and the output feedback, and continuously updates the generative adversarial network model and the causal relationship graph under unsupervised or weakly supervised conditions, so that the system gradually adapts to new public opinion dynamics. The opportunity early warning output module outputs public opinion analysis reports and business opportunity mining results, including causal chain links, trend predictions, and corresponding decision recommendations or early warning prompts, based on the analysis results of the above modules. 2.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The cross-cultural emotion recognition module comprises a multi-language text encoding submodule and an emotion prediction submodule. The output of the emotion prediction submodule is trained by aligning with the emotion features extracted from the electroencephalogram signals. Specifically, for training samples containing electroencephalogram annotations, the error between the text emotion prediction value and the corresponding electroencephalogram emotion features is minimized, so that the deep learning model output of the emotion representation of the text in different languages is consistent with the actual emotion response of different cultural groups. 3.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The causal graph neural network analysis module uses a hierarchical graph neural network structure to process causal graph nodes by granularity and considers different time lags to identify short-term and long-term causal relationships. The causal graph neural network analysis module learns the causal graph structure through causal effect tests and sparse regularization constraints to distinguish between true causal relationships and only correlated relationships. 4.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The emotional counterfactual simulation module comprises: a scenario setting unit for defining at least two different development scenarios; a simulation calculation unit for iteratively calculating the changing trajectories of public opinion emotional indicators and related market indicators over time in each scenario based on the current causal graph; a result comparison unit for comparing the simulation results of different scenarios to evaluate the potential impact of a specific intervention or event on the market, outputting key difference indicators and recommended decision points. 5.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The cross-context nested causal chain opportunity discovery module judges whether the chain crosses different languages, regions or industries by identifying the context attribute differences of different nodes in the causal chain; for cross-context causal chains, further evaluate their reliability and impact range, output high reliability and high impact chains as potential business opportunities; the cross-context nested causal chain opportunity discovery module also combines the scenario analysis of the emotional counterfactual simulation module to determine whether there is an intervention point that can significantly improve business opportunity income by changing a certain antecedent scenario, and gives corresponding action suggestions. 6.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The generative adversarial network of the abnormal emotional group detection module includes two parts of a discriminator and a generator; the discriminator is used to learn the distribution of normal public opinion emotional patterns, and outputs the probability that the input actual public opinion data belongs to normal or abnormal; the generator generates simulated public opinion data samples with noise and condition vectors as input, and is trained against the discriminator; in the detection stage, if the discriminator determines that the input data is abnormal, the public opinion group to which the data belongs is marked as an abnormal emotional group; the module sends the feature information of the abnormal group to the cross-context nested causal chain opportunity discovery module and the opportunity warning output module for corresponding risk prompts and in-depth analysis. 7.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The causal graph generative learning engine uses the difference between the prediction results of the causal relationship graph on market indicators and the actual observation data as a feedback signal, and dynamically adjusts the causal graph using an unsupervised structure learning algorithm; the causal graph generative learning engine further combines the generative adversarial training mechanism, and regards the entire public opinion to market analysis process as a generative model, and evaluates the matching degree with the real situation by the discrimination process, and continuously optimizes the parameters of each functional module. 8.The international market public opinion analysis and business opportunity mining system based on deep learning of claim 1, wherein, The opportunity warning output module provides an interpretable causal chain basis and quantitative evaluation results for each potential business opportunity, including key public opinion events involved in the business opportunity, affected market indicators, expected impact time lag and amplitude, and action plans recommended by the system for the business opportunity; the opportunity warning output module supports displaying the business opportunity and warning information to the user through a visual interface, and allows the user to feed back and confirm or adjust the analysis results, so as to incorporate artificial professional knowledge into the causal graph generative learning engine to further optimize the system performance.

9. A deep learning-based international market public opinion analysis and business opportunity mining method, characterized in that, The application is applied to the international market public opinion analysis and business opportunity mining system based on deep learning in any one of claims 1-8, and specifically comprises the following steps: Step 1: Obtain public opinion text data from multiple countries and languages, and preprocess the data; collect part of the brain electrical signal of the sample in the training stage for emotion model calibration; Step 2: Input the pre-processed public opinion text in each language into the pre-trained cross-cultural emotion recognition model, and output the unified emotion representation and emotion intensity indicators with emotion scales; Step 3: According to the emotion representation and related market historical data, a multi-granularity public opinion-market causal relationship graph is constructed; graph neural networks and causal inference algorithms are used to update the node representation and causal connection of the causal graph, and the implicit causal path in the evolution of public opinion is identified; Step 4: For the identified important public opinion events or links, conduct counterfactual scenario simulation to predict the changes of market indicators under different hypothetical conditions; combined with the causal link structure, potential cross-context supply or demand gap opportunities are explored, and their impact is evaluated; Step 5: Monitor the emotion data stream and use the generative adversarial network to identify abnormal emotion groups; for the detected abnormalities, generate early warning information in a timely manner and include it in the causal analysis range; Step 6: Output the public opinion analysis report and business opportunity mining results for user decision-making reference; at the same time, collect actual market posterior data and user feedback, and feed them back to the causal graph generative learning engine to update the causal graph neural network analysis module and abnormal emotion group detection module, realizing continuous adaptive learning. 10.The international market public opinion analysis and business opportunity mining method based on deep learning according to claim 9, characterized in that, The training of the cross-cultural emotion recognition model in step 2 includes: using samples containing electroencephalogram signals, calculating the mean square error loss between the emotion representation output by the text model and the emotion features extracted from the electroencephalogram, and combining it with the regular sentiment classification loss to minimize it to obtain model parameters that can accurately predict emotions across languages. During inference, the input of public opinion text in any language outputs a vector representation in the common emotion space; The construction of the causal relationship graph in step 3 includes: performing correlation analysis on public opinion emotion indicators and market indicators at different time scales to screen candidate causal relationships with time dependence; using a structured graph neural network to evaluate candidate causal relationships, eliminating false associations and retaining true causal relationships to obtain a multi-level connected causal network; and continuously updating the causal network with new data during system operation to make the causal relationship graph dynamically evolve; The counterfactual scenario simulation in step 4 includes: determining the current state parameters of the target public opinion event; setting the evolution parameters of the event in the future under at least one hypothetical scenario; based on the causal relationship graph, simulating the changes of public opinion and related market indicators step by step from the initial conditions of the hypothetical scenario; record and compare the predicted trajectories of key indicators under different scenarios to obtain scenario differences and key influencing factors for business opportunity evaluation; The abnormal emotion group detection in step 5 includes: inputting real-time public opinion emotion data into a pre-trained generative adversarial network discriminant model to output its probability of belonging to a normal pattern; if the probability value is below a certain threshold, the corresponding public opinion group is determined to be an abnormal emotion group; record the attributes and emotion characteristics of the abnormal group, output early warning information, and further simulate and verify the abnormal pattern through the generative model to improve the confidence of the abnormal detection result; The continuous self-adaptive learning in step 6 comprises: comparing the emotional counterfactual simulation prediction result with the actual market posterior data, calculating the error and adjusting the weight or structure of the corresponding causal edge in the causal graph atlas accordingly; adopting the result of the anomaly detection, incorporating the abnormal public opinion event into the nodes of the causal graph, and trying to establish the causal relationship between the abnormal public opinion event and other nodes; correcting the model output according to the user feedback, adding or deleting the corresponding relationship in the causal graph, so as to continuously improve the accuracy of system analysis and prediction.

Images