Development method of intelligent public opinion analysis and prediction system based on deep learning

Abstract

The invention relates to the technical field, and discloses a development method of an intelligent public opinion analysis and prediction system based on deep learning. Comprising the following steps: a data acquisition step, a data preprocessing step, a feature extraction step, a public opinion analysis model construction step, a public opinion prediction model construction step, a model training and optimization step, a public opinion monitoring and real-time analysis step and a public opinion prediction and early warning step. According to the development method of the intelligent public opinion analysis and prediction system based on deep learning and the related implementation modes, through a series of technical means and optimization measures, the quality and value of public opinion data can be effectively improved, the accuracy and comprehensiveness of public opinion analysis are enhanced, the precision and reliability of public opinion prediction are improved, and the development efficiency is improved. The model training and updating efficiency is optimized, and timely and effective public opinion early warning is realized.

Description

Technical Field

[0001] The present invention relates to the technical field, and specifically to a method for developing an intelligent public opinion analysis and prediction system based on deep learning. Background Art

[0002] In today's information-rich world, public opinion spreads rapidly and has widespread impact. Businesses, governments, and various social organizations all need to understand public opinion trends in a timely and accurate manner. For businesses, public opinion can impact product market image, sales performance, and corporate reputation. Governments need to monitor public opinion to formulate sound policies and maintain social stability. Social organizations also need to adapt their development strategies and public relations tactics based on public opinion. Therefore, public opinion analysis and prediction has become a crucial research area. Summary of the Invention

[0003] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a method for developing an intelligent public opinion analysis and prediction system based on deep learning to solve the above problems.

[0004] (2) Technical solution To achieve the above objectives, the present invention provides the following technical solution: a method for developing an intelligent public opinion analysis and prediction system based on deep learning, comprising the following steps: The data collection step is used to collect public opinion-related data from multiple data sources. The data sources are selected based on a predefined industry weight matrix. The industry weight matrix is ​​determined based on the influence, sensitivity, and timeliness of different industries in the dissemination of public opinion, so as to prioritize the collection of data that is critical to public opinion analysis and prediction; The data preprocessing step involves cleaning, denoising, and preprocessing the collected data to improve data quality. Cleaning includes removing semantically duplicated data based on semantic similarity algorithms, correcting erroneous data due to differences in data source formats, and deleting information unrelated to predefined core public opinion topics. Denoising uses an adaptive filtering algorithm to dynamically adjust filtering parameters based on the data's distribution characteristics to remove noise. Preprocessing also includes word segmentation, part-of-speech tagging, and named entity recognition based on multilingual pre-trained models to adapt to multilingual public opinion data. The feature extraction step uses a deep learning algorithm to extract features from the preprocessed data to obtain a feature vector that can reflect public opinion information. The deep learning algorithm is a hybrid model of a convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM) that integrates an attention mechanism. The attention mechanism dynamically focuses on key feature areas based on the keyword weights in the public opinion data for feature extraction, improving the feature vector's ability to represent public opinion semantic information. The public opinion analysis model construction step uses the extracted feature vectors to construct a public opinion analysis model. This model can classify public opinion data, perform sentiment analysis, and extract topics. Classification uses a multi-classifier ensemble method based on adversarial training, and the discriminator of the generative adversarial network (GAN) and multiple classifiers with different structures work together to improve classification accuracy. Sentiment analysis uses a method that integrates sentiment dictionaries and deep learning semantic understanding to combine prior sentiment knowledge with the dynamic understanding of semantics by the deep learning model to determine the emotional tendency of public opinion information. Topic extraction uses a topic model algorithm that integrates the hierarchical Dirichlet process (HDP) and deep neural network to extract key topics from public opinion data to adapt to topic mining at different levels and granularities. The public opinion prediction model is constructed using deep learning technology based on historical public opinion data and analysis results to predict future public opinion trends. The public opinion prediction model is based on a graph neural network (GNN) structure, which constructs the entity relationships in the public opinion data as graph structure data. The GNN learns the impact of changes in the relationships between entities on public opinion trends. At the same time, it integrates multimodal data, including text, image, and video information, and comprehensively considers the synergistic impact of different modal data on the development of public opinion to improve the accuracy of predictions. The model training and optimization steps use a large amount of labeled data to train the public opinion analysis model and the public opinion prediction model, and optimize them by adjusting the model parameters. The labeled data is obtained using an active learning strategy, which actively selects the most valuable data for manual labeling based on the uncertainty of the model on unlabeled data, reducing the labeling workload. The meta-learning algorithm is introduced to optimize the hyperparameter adjustment during the model training process, quickly converging to the optimal hyperparameter combination based on the different distribution characteristics of public opinion data. The public opinion monitoring and real-time analysis step inputs the real-time collected public opinion data into the trained public opinion analysis model for real-time public opinion monitoring and analysis. In this step, an incremental learning algorithm is used to update the public opinion analysis model in real time, enabling the model to quickly adapt to new public opinion data features. At the same time, a distributed computing framework is used to achieve parallel processing of real-time data, improving analysis efficiency. The public opinion prediction and early warning steps use the public opinion prediction model to predict the future development of public opinion and issue early warning signals according to the preset early warning rules. The early warning rules are based on dynamic threshold settings, and the early warning thresholds are adaptively adjusted according to the historical fluctuation range and current development speed of public opinion. Early warning signals include text messages, emails or system pop-ups, and are pushed in a graded manner according to the urgency and importance of public opinion.

[0005] Preferably, in the data collection step, the data sources include social media platforms, news websites, forums, and blogs, and different collection frequencies are used for different data sources, and the collection frequency is dynamically adjusted according to the information update speed and data volume of the data source.

[0006] Preferably, in the data preprocessing step, when performing a semantic similarity algorithm to remove semantically duplicated data, a hybrid method based on word vectors and semantic role labeling is used to calculate semantic similarity.

[0007] Preferably, in the step of constructing the public opinion analysis model, in the multi-classifier integration method based on adversarial training, classifiers with different structures include improved versions of support vector machines (SVMs), decision trees, and naive Bayes classifiers. The improvement lies in taking the feature vectors extracted by deep learning as input and adjusting the corresponding kernel functions or decision rules.

[0008] Preferably, in the step of constructing the public opinion prediction model, when fusing multimodal data, a multimodal attention mechanism is adopted to dynamically allocate weights according to the contribution of different modal data to the public opinion trend.

[0009] Preferably, in the model training and optimization steps, when using a meta-learning algorithm to optimize hyperparameter adjustment, a model-based meta-learning method is used to use the structural information of the public opinion analysis and prediction model itself as part of the meta-learning for hyperparameter optimization.

[0010] Preferably, in the steps of public opinion monitoring and real-time analysis, when using a distributed computing framework to implement parallel processing of real-time data, a distributed architecture based on containerization technology is adopted to improve the scalability and resource utilization of the system.

[0011] Preferably, in the public opinion prediction and early warning step, when the early warning signal is pushed in a graded manner, different types of public opinions are graded according to user-defined areas of concern and priorities.

[0012] Preferably, it stores a computer program for implementing the method for developing an intelligent public opinion analysis and prediction system based on deep learning as recited in any one of claims 1 to 8.

[0013] Compared with the existing technology, the present invention provides a method for developing an intelligent public opinion analysis and prediction system based on deep learning, which has the following beneficial effects: The present invention's development method and related implementation methods for a deep learning-based intelligent public opinion analysis and prediction system, through a series of technical means and optimization measures, can effectively improve the quality and value of public opinion data, enhance the accuracy and comprehensiveness of public opinion analysis, improve the precision and reliability of public opinion predictions, optimize the efficiency of model training and updating, and achieve timely and effective public opinion warnings. These beneficial effects give the system significant advantages in public opinion monitoring, analysis, and prediction, and can provide strong support for businesses, governments, and other relevant institutions, helping them better understand public opinion trends, respond promptly to public opinion crises, formulate reasonable decisions and strategies, and maintain social stability and the public interest. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the embodiments described are only some embodiments of the invention, not all embodiments. In the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the invention.

[0016] See also Figure 1 , a development method for an intelligent public opinion analysis and prediction system based on deep learning: 1. Data Collection Steps Data source determination and weight matrix application First, a predefined weighting matrix for industry sectors is established. For example, the financial sector is given a higher weight due to its importance in the economic system, sensitivity to market fluctuations, and high timeliness of information. Meanwhile, relatively niche industries with a smaller scope of public opinion dissemination are given lower weights.

[0017] According to this weight matrix, data sources are selected, including social media platforms (such as Weibo, Twitter, etc.), news websites (such as Sina News, Tencent News, etc.), forums (such as Tianya Forum, Baidu Tieba, etc.), and blogs (such as Sina Blog, etc.).

[0018] Acquisition frequency adjustment Set different collection frequencies for different data sources. For social media platforms with fast updates and large data volumes, such as Weibo, data might be collected every 10-15 minutes; whereas for blogs with slower updates, data might be collected every 1-2 hours. The collection frequency will be adjusted dynamically based on the actual update rate and data volume of the data source. For example, if a social media platform experiences an explosive growth of new topics and information within a certain period of time, the collection frequency will be appropriately increased.

[0019] 2. Data Preprocessing Steps Cleaning operation Semantic deduplication When using a semantic similarity-based algorithm to remove semantically duplicated data, a hybrid approach based on word embeddings and semantic role annotations is used to calculate semantic similarity. For example, for sentences in text data, the words are first converted into word embeddings. Then, combined with the semantic role annotation information, the semantic distance between the sentences is calculated. If the semantic distance between two sentences is less than a preset threshold (such as 0.8), they are considered semantically duplicated and one of them is removed.

[0020] Error data correction Correct erroneous data due to format differences in data sources. For example, different news websites may have different date formats, some using "year-month-day" and others using "month / day / year." By writing format conversion rules, these formats can be unified into a standard format.

[0021] Delete irrelevant information Based on the predefined core public opinion topic, irrelevant information is deleted. For example, if the core public opinion topic is about the launch of a new mobile phone, then information related to other mobile phone brands or unrelated products will be deleted.

[0022] Denoising operation Adaptive filtering algorithms are used to remove noise. First, the data distribution characteristics are analyzed. For example, for numerical data in public opinion data (such as public opinion popularity values), if the data exhibits a normal distribution, the filtering parameters are dynamically adjusted based on the normal distribution parameters (mean and standard deviation). Data points that deviate from the mean by a certain multiple of the standard deviation (e.g., 3 times the standard deviation) are considered noise points and filtered out.

[0023] Text data preprocessing Perform word segmentation, part-of-speech tagging, and named entity recognition on text data using multilingual pre-trained models (such as the multilingual version of BERT). For example, if a public opinion document contains multiple languages ​​(such as Chinese and English), it is input into a multilingual pre-trained model. The model will output the word segmentation results, part-of-speech tags (such as nouns and verbs), and named entities (such as names of people, places, and organizations) for each word for subsequent processing.

[0024] 3. Feature extraction steps Hybrid deep learning model construction Construct a hybrid model that integrates a convolutional neural network (CNN) with an attention mechanism and a bidirectional long short-term memory network (Bi-LSTM).

[0025] When processing public opinion data, the attention mechanism dynamically focuses on key feature areas based on the keyword weights within the data for feature extraction. For example, when analyzing public opinion regarding a company's product quality issues, keywords such as "product quality" and "recall" have higher weights. The attention mechanism will focus more on the text regions containing these keywords, thereby improving the feature vector's ability to represent the semantic information of public opinion. By inputting preprocessed public opinion data into this hybrid model, a feature vector that reflects public opinion information is obtained.

[0026] 4. Steps to build a public opinion analysis model Classification Operation In a multi-classifier ensemble method based on adversarial training, a generative adversarial network (GAN) discriminator is used in conjunction with multiple classifiers of different structures to improve classification accuracy. These classifiers include improved versions of support vector machines (SVMs), decision trees, and naive Bayes classifiers.

[0027] For the support vector machine (SVM), the feature vector extracted by deep learning is used as input, and the corresponding kernel function is adjusted (such as adjusting the linear kernel function to the radial basis kernel function); for the decision tree, the decision rules are adjusted according to the feature vector, for example, the basis for dividing the branch nodes is re-determined according to the importance of the feature; for the improved version of the naive Bayes classifier, the feature vector is also used as input, and the relevant parameters of the probability calculation are adjusted.

[0028] Sentiment analysis operations The sentiment leaning of public opinion information is determined by integrating a sentiment lexicon with deep learning semantic understanding. For example, a sentiment lexicon is first constructed containing common sentiment words (such as "happy," "sad," and "angry") and their sentiment polarity (positive, negative, and neutral). The public opinion text is then fed into a deep learning model (such as a recurrent neural network based on an attention mechanism). The model dynamically understands the text's semantics and incorporates prior sentiment knowledge from the sentiment lexicon. If the text contains a high number of negative words from the sentiment lexicon, and the deep learning model's understanding of the semantics also leans towards negativity, the public opinion information is determined to have a negative sentiment leaning.

[0029] Topic extraction operation A topic modeling algorithm that combines a hierarchical Dirichlet process (HDP) and a deep neural network is used to extract key topics from public opinion data. For example, for public opinion data containing multiple subtopics, the HDP can automatically determine the hierarchical structure of the topics, while a deep neural network (such as a multilayer perceptron) can further explore the semantics of the topics, thus adapting to topic mining at different levels and granularities.

[0030] 5. Steps for Building a Public Opinion Prediction Model Graph neural network structure construction The entity relationships in public opinion data are constructed as graph-structured data. For example, in public opinion about competition in a certain industry, each competing company is regarded as a node, and the competitive relationship, cooperative relationship, etc. between companies are regarded as edges.

[0031] Graph neural networks (GNNs) are used to learn how changes in relationships between entities affect public opinion trends.

[0032] Multimodal data fusion When fusing multimodal data (including text, images, and video information), a multimodal attention mechanism is employed. For example, when analyzing public opinion about a particular event, if key images in an image or video (such as footage from the incident scene) have a significant impact on public opinion trends, the multimodal attention mechanism will dynamically assign a higher weight based on their contribution to the public opinion trend. This comprehensively considers the synergistic impact of different modal data on the development of public opinion to improve prediction accuracy.

[0033] 6. Model training and optimization steps Active learning strategy to obtain labeled data Active learning strategies are used to obtain labeled data. Unlabeled data is first fed into the public opinion analysis and prediction models. Based on the model's uncertainty about the unlabeled data (e.g., high entropy values ​​in the prediction results), the most valuable data is proactively selected for manual labeling. For example, if the model's classification results for a piece of public opinion data show small differences in probability across multiple categories, indicating uncertainty about the data, this data is prioritized for manual labeling, thus reducing the labeling workload.

[0034] Meta-learning algorithm optimization hyperparameter adjustment Utilize model-based meta-learning to optimize hyperparameter tuning. The structural information of the public opinion analysis and prediction model (such as the number of layers and neurons) is incorporated into the meta-learning process. For example, based on the distribution characteristics of public opinion data (e.g., normal distribution, skewed distribution, etc.), the meta-learning algorithm analyzes the model structure and data characteristics to quickly converge on the optimal hyperparameter combination.

[0035] 7. Steps for Public Opinion Monitoring and Real-time Analysis Incremental learning algorithm updates the model Real-time public opinion data is fed into a trained public opinion analysis model for real-time public opinion monitoring and analysis. An incremental learning algorithm is used to update the public opinion analysis model in real time. For example, when new public opinion data contains new vocabulary or new topic patterns, the incremental learning algorithm adjusts some of the model's parameters without retraining the entire model, allowing the model to quickly adapt to the new public opinion data characteristics.

[0036] Distributed computing framework parallel processing Utilize a distributed architecture based on containerization technologies (such as Docker) to achieve parallel processing of real-time data. For example, real-time public opinion data can be divided into different containers according to certain rules (such as data source or topic type). Each container processes data in parallel on different computing nodes, improving system scalability and resource utilization.

[0037] 8. Public Opinion Prediction and Early Warning Steps Dynamic threshold setting warning rules The warning threshold is adaptively adjusted based on the historical fluctuation range and current development speed of public opinion. For example, if a public opinion topic has had a small fluctuation range in the past, but the current public opinion heat value rises rapidly and exceeds the dynamic threshold calculated based on historical data, an alert will be triggered.

[0038] Early warning signal grading push Different types of public opinion are graded based on user-defined areas of concern and priorities. For example, if a user sets public opinion related to their own company as the highest priority, any negative public opinion about the company will be prioritized via SMS, even if the warning level is low. On the other hand, general industry public opinion, if the warning level is low, may be sent via email. Warning signals may be sent via SMS, email, or system pop-up windows.

[0039] IX. Implementation Based on Computer-Readable Storage Medium Program storage and execution The computer program used to implement the above-mentioned deep learning-based intelligent public opinion analysis and prediction system development method is stored in a computer-readable storage medium (such as a hard disk, solid-state drive, USB flash drive, etc.).

[0040] When the system needs to be run, the computer program is read from the storage medium and run on the corresponding computing device (such as a server, personal computer, etc.), thereby realizing the functions of the entire intelligent public opinion analysis and prediction system.

Claims

1. A method for developing an intelligent public opinion analysis and prediction system based on deep learning, characterized in that: The following steps are involved: A data collection step is used to collect public opinion-related data from multiple data sources, wherein the selection of data sources is based on a predefined industry weight matrix. The industry weight matrix is ​​determined based on the influence, sensitivity, and timeliness of different industries in the dissemination of public opinion, so as to prioritize the collection of data with key value for public opinion analysis and prediction; The data preprocessing step involves cleaning, denoising, and preprocessing the collected data to improve data quality. Cleaning includes removing semantically duplicated data based on semantic similarity algorithms, correcting erroneous data due to differences in data source formats, and deleting information unrelated to predefined core public opinion topics. Denoising uses an adaptive filtering algorithm to dynamically adjust filtering parameters based on the data's distribution characteristics to remove noise. Preprocessing also includes word segmentation, part-of-speech tagging, and named entity recognition based on multilingual pre-trained models to adapt to multilingual public opinion data. A feature extraction step uses a deep learning algorithm to extract features from the preprocessed data to obtain a feature vector that can reflect public opinion information. The deep learning algorithm is a hybrid model of a convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM) that integrates an attention mechanism. The attention mechanism dynamically focuses on key feature areas based on the keyword weights in the public opinion data for feature extraction, thereby improving the feature vector's ability to represent public opinion semantic information. The public opinion analysis model construction step uses the extracted feature vectors to construct a public opinion analysis model. This model can classify public opinion data, perform sentiment analysis, and extract topics. Classification uses a multi-classifier ensemble method based on adversarial training, and the discriminator of the generative adversarial network (GAN) and multiple classifiers with different structures work together to improve classification accuracy. Sentiment analysis uses a method that integrates sentiment dictionaries and deep learning semantic understanding to combine prior sentiment knowledge with the dynamic understanding of semantics by the deep learning model to determine the emotional tendency of public opinion information. Topic extraction uses a topic model algorithm that integrates the hierarchical Dirichlet process (HDP) and deep neural network to extract key topics from public opinion data to adapt to topic mining at different levels and granularities. The step of constructing a public opinion prediction model uses deep learning technology to construct a public opinion prediction model based on historical public opinion data and analysis results to predict future public opinion development trends. The public opinion prediction model is based on a graph neural network (GNN) structure, which constructs the entity relationships in the public opinion data as graph structure data. The graph neural network learns the impact of changes in the relationships between entities on public opinion trends. At the same time, it integrates multimodal data, including text, image and video information, and comprehensively considers the synergistic impact of different modal data on the development of public opinion to improve the accuracy of prediction; The model training and optimization steps use a large amount of labeled data to train the public opinion analysis model and the public opinion prediction model, and optimize them by adjusting the model parameters. The labeled data is obtained using an active learning strategy, which actively selects the most valuable data for manual labeling based on the uncertainty of the model on unlabeled data, reducing the labeling workload. The meta-learning algorithm is introduced to optimize the hyperparameter adjustment during the model training process, quickly converging to the optimal hyperparameter combination based on the different distribution characteristics of public opinion data. The public opinion monitoring and real-time analysis step inputs the real-time collected public opinion data into the trained public opinion analysis model for real-time public opinion monitoring and analysis. In this step, an incremental learning algorithm is used to update the public opinion analysis model in real time, enabling the model to quickly adapt to new public opinion data features. At the same time, a distributed computing framework is used to achieve parallel processing of real-time data, improving analysis efficiency. The public opinion prediction and early warning steps use the public opinion prediction model to predict the future development of public opinion and issue early warning signals according to the preset early warning rules. The early warning rules are based on dynamic threshold settings, and the early warning thresholds are adaptively adjusted according to the historical fluctuation range and current development speed of public opinion. Early warning signals include text messages, emails or system pop-ups, and are pushed in a graded manner according to the urgency and importance of public opinion.

2. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the data collection step, the data sources include social media platforms, news websites, forums, and blogs, and different collection frequencies are used for different data sources. The collection frequency is dynamically adjusted according to the information update speed and data volume of the data source.

3. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the data preprocessing step, when performing a semantic similarity algorithm to remove semantically duplicated data, a hybrid method based on word vectors and semantic role labeling is used to calculate semantic similarity.

4. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the step of constructing the public opinion analysis model, in the multi-classifier integration method based on adversarial training, classifiers with different structures include improved versions of support vector machines (SVMs), decision trees, and naive Bayes classifiers. The improvement lies in taking the feature vectors extracted by deep learning as input and adjusting the corresponding kernel functions or decision rules.

5. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the step of constructing the public opinion prediction model, when fusing multimodal data, a multimodal attention mechanism is adopted to dynamically allocate weights according to the contribution of different modal data to the public opinion trend.

6. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the model training and optimization steps, when using the meta-learning algorithm to optimize hyperparameter adjustment, a model-based meta-learning method is used to optimize the hyperparameters by taking the structural information of the public opinion analysis and prediction model itself as part of the meta-learning.

7. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the public opinion monitoring and real-time analysis steps, when using a distributed computing framework to achieve parallel processing of real-time data, a distributed architecture based on containerization technology is adopted to improve the scalability and resource utilization of the system.

8. The method for developing an intelligent public opinion analysis and prediction system based on deep learning according to claim 1, characterized in that: In the public opinion prediction and early warning step, when the early warning signal is pushed in a graded manner, different types of public opinions are graded according to the user-defined areas of concern and priorities.

9. A computer-readable storage medium storing a computer program for implementing the method for developing an intelligent public opinion analysis and prediction system based on deep learning as described in any one of claims 1 to 8.

Images