Automobile public opinion analysis method
By combining sentiment classification models and latent Dirichlet assignment models with text network analysis, this approach addresses the lack of refinement in traditional sentiment analysis methods, enabling a deeper understanding of public opinion and more precise strategy formulation.
Patent Information
- Application Number
- CN202511779583.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-02-27
AI Technical Summary
Existing methods for analyzing public opinion in the automotive industry are insufficient to reveal the reasons for changes in public opinion, and traditional deep learning technologies struggle to clarify the inherent logical relationships between topics, resulting in insufficiently refined strategy formulation.
By combining a sentiment classification model with a latent Dirichlet assignment model and text network analysis, and by acquiring automotive public opinion data, we can perform sentiment classification, topic mining, and co-occurrence matrix construction to generate a text network graph and reveal the results of public opinion analysis.
Capture the inherent connections between various themes and keywords to help automakers grasp public perception of their brands in real time, develop targeted strategies, and enhance brand image and market performance.
Smart Images
Figure CN121580131A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of public opinion analysis technology, and in particular to a method for analyzing public opinion related to automobiles. Background Technology
[0002] With the rapid development of internet technology, online media has become a major channel for information dissemination. The emotions contained in public opinion not only reveal the public's attitude towards specific events but also reflect the positive or negative tendencies of public opinion, thus influencing the progress of events and the response strategies of decision-makers. Particularly in the automotive industry, public opinions and views on car products have a direct impact on automakers' brand building and market performance. Positive public opinion can strengthen brand image and expand market share; conversely, negative public opinion, if mishandled, can seriously damage a company's reputation and sales performance. Therefore, for automotive companies, implementing effective public opinion monitoring and analysis is crucial.
[0003] Current methods for analyzing public opinion in the automotive industry rely on traditional deep learning techniques to identify and classify sentiment in user comments, and use topic models to uncover core topics of public concern, thereby formulating targeted strategies for refined control of public opinion. Although topic analysis can identify topics of public concern, the model may assign adjacent words to different topics, making it difficult to clarify the inherent logical relationships between topics, and thus failing to reveal the reasons for changes in public opinion at a deeper level. Summary of the Invention
[0004] In view of the above, the present invention aims to provide a method for analyzing public opinion on automobiles to solve the aforementioned technical problems.
[0005] The technical solution adopted in this invention is as follows:
[0006] This invention provides a method for analyzing public opinion on automobiles, including:
[0007] Obtain automotive public opinion data;
[0008] The automotive public opinion data is input into the trained sentiment classification model to obtain the sentiment classification results of the automotive public opinion information.
[0009] The potential Dirichlet assignment model was used to mine topics in automotive public opinion data and obtain topic keywords.
[0010] A co-occurrence matrix is constructed based on topic keywords, a text network graph is generated and analyzed, and the text network analysis results are obtained.
[0011] Based on the sentiment classification results, topic keywords, and text network analysis results, the public opinion analysis results are obtained.
[0012] Optionally, obtain automotive public opinion data, including:
[0013] By scraping text reviews of vehicles from social media platforms, we can obtain car review data.
[0014] The car review data is preprocessed to obtain car public opinion data; the car public opinion data includes...
[0015] Comment content, number of comments, user ID, comment time, vehicle model configuration, and comment IP address.
[0016] Optionally, the car review data may be preprocessed, including:
[0017] Remove duplicate characters, spelling errors, tags, images, links, and special characters;
[0018] Delete comments with fewer than five characters;
[0019] Remove stop words, expand abbreviations, and unify similar words.
[0020] Optional methods for analyzing automotive public opinion also include:
[0021] By jointly analyzing vehicle configurations and the IP addresses of reviews, we can obtain vehicle purchase information results.
[0022] Optionally, the automotive public opinion data is input into a trained sentiment classification model to obtain sentiment classification results for automotive public opinion information, including:
[0023] Text comments are converted into fixed-dimensional text vectors using word embeddings;
[0024] Input the text vector into the trained sentiment classification model, and output the probability vector in the context of the text vector;
[0025] Based on the probability vector, the sentiment classification results of automotive public opinion information are obtained.
[0026] Optionally, training the sentiment classification model includes:
[0027] Sentiment polarity labeling is performed on automotive public opinion data to obtain an automotive public opinion dataset, where sentiment polarity includes positive, neutral, and negative.
[0028] The automotive public opinion dataset is divided into a training set, a validation set, and a test set;
[0029] We construct a sentiment classification model that integrates word embeddings and multiple deep learning architectures, train the model using a training set, and employ hyperparameter optimization and early stopping strategies during the training process.
[0030] The model's hyperparameters were tuned using the validation set, and the model's performance was evaluated using the test set to obtain the evaluation results.
[0031] The model with the best evaluation results was selected as the sentiment classification model.
[0032] Optionally, the deep learning architecture includes at least one of the following:
[0033] Deep neural networks;
[0034] Convolutional Neural Network;
[0035] Long Short-Term Memory (LSTM) networks;
[0036] Bidirectional long short-term memory network.
[0037] Optionally, the Latent Dirichlet Allocation Model can be used to mine topics in automotive sentiment data to obtain topic keywords, including:
[0038] Mining topic sets from automotive sentiment data using the potential Dirichlet assignment model;
[0039] The number of topics in the topic set is determined by the topic consistency index and the confusion index, and the topics are labeled to obtain topic keywords.
[0040] Optionally, a co-occurrence matrix is constructed based on topic keywords, a text network graph is generated and analyzed, and the text network analysis results are obtained, including:
[0041] Construct a keyword co-occurrence matrix;
[0042] The text network graph is generated based on the co-occurrence matrix, and the degree centrality, betweenness centrality, and proximity centrality of the nodes are calculated to obtain the text network analysis results.
[0043] Optionally, in the text network graph, the edge thickness represents the probability of topic co-occurrence in the automotive public opinion data, and the node color is determined based on the Beta value of the topic keyword in the topic set.
[0044] The above-described solution of the present invention has at least the following beneficial effects:
[0045] The above-described solution of the present invention involves: acquiring automotive public opinion data; inputting the automotive public opinion data into a trained sentiment classification model to obtain sentiment classification results for automotive public opinion information; using a latent Dirichlet allocation model to mine the themes in the automotive public opinion data to obtain theme keywords; constructing a co-occurrence matrix based on the theme keywords, generating a text network graph and analyzing it to obtain text network analysis results; and obtaining public opinion analysis results based on the sentiment classification results, theme keywords, and text network analysis results. The solution of the present invention captures the inherent relationship between various themes and keywords, helping automakers to grasp public perception of their brands in real time and formulate targeted strategies. Attached Figure Description
[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings, wherein:
[0047] Figure 1 A flowchart of the automotive public opinion analysis method provided in an embodiment of the present invention.
[0048] Figure 2 This is a schematic diagram illustrating the probability values of public attention to a certain car model from a car manufacturer, provided as an embodiment of the present invention.
[0049] Figure 3 This is a schematic diagram of a text network representing a topic of interest, provided as an embodiment of the present invention. Detailed Implementation
[0050] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0051] This invention proposes an embodiment of a method for analyzing public opinion on automobiles, specifically, as follows: Figure 1 As shown, it includes:
[0052] Step 11, obtain automotive public opinion data;
[0053] In this embodiment, text reviews about vehicles are crawled from automotive media platforms to obtain automotive review data. Data preprocessing is then performed, including: removing duplicate characters and spelling errors using regular expressions; filtering tags, images, links, and special symbols; removing short reviews of less than five characters to maintain contextual semantic relevance; removing irrelevant Chinese stop words to improve model accuracy; expanding the full expression of abbreviations; unifying similar words to simplify the corpus and improve word embedding generation. The preprocessed data yields automotive public opinion data, including review content, number of reviews, URL, user ID, review time, vehicle model configuration, and review IP address.
[0054] Step 12: Input the automotive public opinion data into the trained sentiment classification model to obtain the sentiment classification results of the automotive public opinion information;
[0055] In this embodiment, a deep learning model is first trained and selected. Sentiment polarity is labeled on public statements to obtain an automotive public opinion dataset, where sentiment polarity includes positive (labeled "1"), neutral (labeled "0"), and negative (labeled "-1"). The automotive public opinion dataset is then divided into a training set (80%), a validation set (10%), and a test set (10%). The training set is used to develop a deep learning-based sentiment classification model, the validation set is used for unbiased evaluation of the trained model during hyperparameter fine-tuning, and the performance evaluation of the classification model is performed on the test set.
[0056] To construct a sentiment classification model, this embodiment combines word embedding (Word2Vec) with four deep learning models: Deep Neural Network (DNN), Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and Bidirectional Long Short-Term Memory Network (Bi-LSTM) for the sentiment classification task of text comments, given the excellent accuracy of deep neural network models in various natural language processing tasks.
[0057] To improve the performance of neural network models on sentiment classification tasks, this embodiment introduces the Skip-gram architecture of Word2Vec vector embedding technology, and based on a two-layer neural network architecture, processes text input and generates a fixed-dimensional text vector representation for each word.
[0058] The Skip-gram model works as follows: each word in the corpus is converted into an X-dimensional one-hot encoded vector, which serves as the input to the Skip-gram model. The number of neurons in the input layer and hidden layer corresponds to the size of the vocabulary and the dimension of the word vectors, respectively. The weight matrix connecting the input layer and hidden layer is... Let X represent the size of the vocabulary and Y represent the dimension of the hidden layer. Each row of the matrix corresponds to a Y-dimensional vector representation of a word. The weight matrix between the hidden layer and the output layer is denoted as... The network is responsible for calculating the score of each word in the vocabulary given the input word. The network ultimately outputs a probability vector showing the probability of each word appearing in the context of the input word. Through backpropagation, the model updates the weights based on the error of the loss function, learning the relationships between words in the vocabulary.
[0059] The initial stage of a degree neural network employs an embedding layer, whose core task is to generate a set of continuous and discriminative vectors for words in a text dataset. This layer is defined by four main parameters.
[0060] The first parameter is the vocabulary size, which is equal to the total number of distinct words in the dataset;
[0061] The second parameter defines the dimension of the embedding vector space, that is, the dimension to which the words will be mapped;
[0062] The third parameter, denoted as K, refers to the maximum length of the input sequence. This parameter is determined based on the percentile distribution of the number of words in the comments in the dataset. For comments with fewer than K words, zero-padding is added to expand them to K; while comments with more than K words need to be truncated to fit the length K.
[0063] The fourth parameter relates to the embedding matrix, which is constructed using the pre-trained word embedding model Word2Vec, assigning a p-dimensional vector to each unique word in the dataset and integrating these vectors into a K×p matrix.
[0064] During model training, the goal of the embedding layer is to find the optimal word vector mapping for each segmented word in the comment to match the preset output dimension. After sufficient training, the embedding matrix is able to capture subtle multidimensional differences between words with similar contextual meanings. Finally, the output of the embedding layer will be used as input to deep neural network models (DNN, CNN, LSTM, and Bi-LSTM) specifically designed for sentiment analysis.
[0065] In the field of deep learning, the choice of hyperparameters has a significant impact on the accuracy and stability of models. This invention employs a grid search strategy to optimize several key hyperparameters, including the number of filters and filter size in convolutional neural networks (CNNs) and the number of units in long short-term memory networks (LSTMs). This strategy involves constructing a grid containing all potential hyperparameter values and evaluating the cross-validation performance of the model under each hyperparameter configuration.
[0066] To improve the model's predictive performance and reduce the risk of overfitting during training, a k(5)-fold cross-validation method is used to train the deep learning model. The specific steps are as follows:
[0067] The entire dataset is evenly divided into k(5) subsets for training and model evaluation; the model is trained using (k-1) subsets and its performance is evaluated on the k-th subset that was not used for training; the above training and evaluation steps are repeated to ensure that each subset is used as a test set; the final accuracy of the model is obtained by averaging the accuracy obtained on all different test subsets.
[0068] This embodiment employs early stopping to determine the appropriate training period for the model. The training process is terminated when the improvement in accuracy stabilizes, meaning that the model performance shows almost no significant improvement in consecutive iterations.
[0069] The classification performance of the models is measured and compared based on four key metrics: accuracy, precision, recall, and F-measure.
[0070] ;
[0071] ;
[0072] ;
[0073] ;
[0074] Wherein, TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples.
[0075] By comparing the classification performance evaluation results of DNN, CNN, LSTM and Bi-LSTM models, the model with the best sentiment classification performance was selected as the sentiment classification model.
[0076] Furthermore, the fixed-dimensional text vector, transformed through word embedding of text comments, is input into the trained sentiment classification model, which outputs a probability vector in the context of the text vector; based on the probability vector, the sentiment classification result of the automotive public opinion information is obtained.
[0077] Step 13: Use the latent Dirichlet assignment model to mine topics in the automotive public opinion data and obtain topic keywords;
[0078] In this embodiment, a topic model is used to mine public interest in automotive products. The topic model is a Latent Dirichlet Allocation (LDA) model, an unsupervised model for probabilistically generating hidden topics in text data. Its workflow is as follows:
[0079] 1) Topic initialization: Randomly assign a topic to each comment;
[0080] 2) Topic content initialization: For identified topics, words are randomly selected from the vocabulary list to initialize the vocabulary distribution of these topics;
[0081] 3) Topic probability calculation: For each word in each comment, calculate the probability that the word belongs to each topic based on the current topic assignment and word distribution;
[0082] 4) Topic Reassignment: Based on the calculated probabilities, the topics of the vocabulary are reassigned. This iterative process will continue to optimize until the model converges.
[0083] The construction of an LDA topic model includes the following steps:
[0084] S51: Estimate the topic distribution and word distribution under each topic for each document, and determine a reasonable number of topics based on the topic consistency index and the perplexity index;
[0085] S52: The LDA model is an unsupervised learning method that uses experts to manually label topics based on the top 10 keywords with the highest probabilities. In one embodiment, the public has six topics of interest regarding a particular car brand's models, such as... Figure 2 As shown, the Beta value represents the probability of a keyword appearing in each topic.
[0086] The rules for topic coding are as follows. Taking Topic 1, Driving Experience, as an example, the keyword "driving" appears more frequently. Other keywords, such as "jumping," "comfort," "handling," and "seat," are also semantically related to driving experience. Therefore, this topic is tagged as "Driving Experience." Similarly, other topics are tagged based on frequently occurring keywords.
[0087] Step 14: Construct a co-occurrence matrix based on topic keywords, generate a text network graph and analyze it to obtain the text network analysis results;
[0088] In this embodiment, a co-occurrence matrix is constructed based on LDA keywords and their co-occurrence relationships, which is mapped to a co-word network to reveal the internal connections between topics. Degree centrality, betweenness centrality, and proximity centrality are three important metrics for evaluating nodes in the text network.
[0089] Degree centrality measures the number of connections a node has in a network, i.e., the node's degree. A higher degree indicates higher degree centrality, meaning the node is more important in the network. The formula for calculating the degree centrality of a node is as follows:
[0090] ;
[0091] in, N is the number of edges connected to node m, and N-1 is the number of edges that connect node m to all other nodes.
[0092] Betweenness centrality measures how well a node connects to other nodes in a network, establishing the shortest paths between them. The formula for calculating the betweenness centrality of a node is as follows:
[0093] ;
[0094] in, This represents the number of paths that pass through node m and are the shortest paths. This represents the number of shortest paths connecting s and t.
[0095] Closeness centrality measures the average distance between a node and other nodes; the shorter the distance, the higher the closeness centrality. The formula for calculating the closeness centrality of a node is as follows:
[0096] ;
[0097] ;
[0098] in, To approximate centrality, Let m be the average distance from node m to the other nodes, n be the number of the other nodes, and N be the total number of nodes.
[0099] Text network analysis specifically includes the following steps:
[0100] The high-frequency words and line features of the text data were extracted using the ROST semantic network generation tool, and a co-occurrence matrix of high-frequency words was constructed.
[0101] Manually check the threshold of keywords for each topic in LDA, determine the threshold for assigning comments to topics to be 0.015, and then select keywords with a threshold greater than 0.015;
[0102] Keywords with a threshold of 0.015 or higher in LDA are selected from the high-frequency word co-occurrence matrix to ensure semantic connection between topics and comments, forming a keyword co-occurrence matrix;
[0103] The co-occurrence matrix was imported into Gephi to create a network diagram, revealing the text's themes and keywords.
[0104] The thickness of edges in a text network is positively correlated with the probability of coexistence of topics in comments; that is, the thicker the edge, the higher the probability of coexistence of topics in the comments.
[0105] In one specific embodiment, keywords may exist in different topics, and the node color of the keywords is mainly determined based on Beta values. In one embodiment, the text network of public interest in a particular car brand's model is as follows: Figure 3 As shown, the keyword "brake" belongs to both Topic 1 (Driving Experience) and Topic 2 (Intelligent Driving). The Beta value for "brake" in Topic 1 (Driving Experience) is 0.022, while the Beta value in Topic 2 (Intelligent Driving) is 0.018. Therefore, the keyword "brake" is classified under Topic 1 (Driving Experience).
[0106] Step 15: Based on the sentiment classification results, topic keywords, and text network analysis results, obtain the public opinion analysis results.
[0107] In this embodiment, based on sentiment classification results, topic keywords, and text network analysis results, the public's views on car companies' products and the reasons for changes in these views are revealed, providing insights for car companies' public opinion management and marketing.
[0108] In an optional embodiment of the present invention, the automotive public opinion analysis method further includes: performing joint analysis of vehicle configuration and comment IP addresses to obtain vehicle purchase information results.
[0109] In this embodiment, vehicle purchase user information (drive type, number of seats and model) is integrated, and maps of various regions are imported according to the user's IP information format. The vehicle purchase information and IP address are analyzed together to realize visualized vehicle purchase information of users in various regions.
[0110] In summary, the automotive public opinion analysis method described in the above embodiments of the present invention employs word embedding technology to map words to a low-dimensional vector space to capture semantic connections and contextual information between words. By integrating a deep learning architecture, the model's ability to process natural language is enhanced. Simultaneously, by combining topic analysis and text network technology, this not only solves the problem that traditional topic models may classify adjacent words into different topics, avoiding topic overlap, but also deeply explores the intrinsic connections between topics and keywords in the text, providing new strategies for public opinion monitoring and marketing in the automotive industry. The fusion of topic analysis and text networks provides in-depth insights for automakers' public opinion analysis, helping them better understand public opinion, respond promptly to market changes, and adopt effective strategies to maintain and enhance their brand image.
[0111] In this invention, when directional terms are mentioned, they are relative concepts based on the embodiments. Furthermore, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.
[0112] The above description of the structure, features, and effects of the present invention is based on the embodiments shown in the figures. However, the above are only preferred embodiments of the present invention. It should be noted that the technical features involved in the above embodiments and their preferred methods can be reasonably combined and matched by those skilled in the art to form a variety of equivalent solutions without departing from or changing the design concept and technical effects of the present invention. Therefore, the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.
Claims
1. A method for analyzing public opinion on automobiles, characterized in that, include: Obtain automotive public opinion data; The automotive public opinion data is input into the trained sentiment classification model to obtain the sentiment classification results of the automotive public opinion information. The potential Dirichlet assignment model was used to mine topics in automotive public opinion data and obtain topic keywords. A co-occurrence matrix is constructed based on topic keywords, a text network graph is generated and analyzed, and the text network analysis results are obtained. Based on the sentiment classification results, topic keywords, and text network analysis results, the public opinion analysis results are obtained.
2. The automotive public opinion analysis method according to claim 1, characterized in that, Obtain automotive public opinion data, including: By scraping text reviews of vehicles from social media platforms, we can obtain car review data. The car review data is preprocessed to obtain car public opinion data; the car public opinion data includes... Comment content, number of comments, user ID, comment time, vehicle model configuration, and comment IP address.
3. The automotive public opinion analysis method according to claim 2, characterized in that, The car review data is preprocessed, including: Remove duplicate characters, spelling errors, tags, images, links, and special characters; Delete comments with fewer than five characters; Remove stop words, expand abbreviations, and unify similar words.
4. The automotive public opinion analysis method according to claim 2, characterized in that, Also includes: By jointly analyzing vehicle configurations and the IP addresses of reviews, we can obtain vehicle purchase information results.
5. The automotive public opinion analysis method according to claim 1, characterized in that, The automotive public opinion data is input into a trained sentiment classification model to obtain sentiment classification results for automotive public opinion information, including: Text comments are converted into fixed-dimensional text vectors using word embeddings; Input the text vector into the trained sentiment classification model, and output the probability vector in the context of the text vector; Based on the probability vector, the sentiment classification results of automotive public opinion information are obtained.
6. The automotive public opinion analysis method according to claim 1, characterized in that, Training the sentiment classification model includes: Sentiment polarity labeling is performed on automotive public opinion data to obtain an automotive public opinion dataset, where sentiment polarity includes positive, neutral, and negative. The automotive public opinion dataset is divided into a training set, a validation set, and a test set; We construct a sentiment classification model that integrates word embeddings and multiple deep learning architectures, train the model using a training set, and employ hyperparameter optimization and early stopping strategies during the training process. The model's hyperparameters were tuned using the validation set, and the model's performance was evaluated using the test set to obtain the evaluation results. The model with the best evaluation results was selected as the sentiment classification model.
7. The automotive public opinion analysis method according to claim 6, characterized in that, The deep learning architecture includes at least one of the following: Deep neural networks; Convolutional Neural Network; Long Short-Term Memory (LSTM) networks; Bidirectional long short-term memory network.
8. The automotive public opinion analysis method according to claim 1, characterized in that, The Latent Dirichlet Allocation Model was used to mine topics in automotive public opinion data, resulting in topic keywords, including: Mining topic sets from automotive sentiment data using the potential Dirichlet assignment model; The number of topics in the topic set is determined by the topic consistency index and the confusion index, and the topics are labeled to obtain topic keywords.
9. The automotive public opinion analysis method according to claim 1, characterized in that, A co-occurrence matrix is constructed based on topic keywords, a text network graph is generated and analyzed, and the text network analysis results are obtained, including: Construct a keyword co-occurrence matrix; The text network graph is generated based on the co-occurrence matrix, and the degree centrality, betweenness centrality, and proximity centrality of the nodes are calculated to obtain the text network analysis results.
10. The automotive public opinion analysis method according to claim 9, characterized in that, In the text network graph, the edge thickness represents the probability of topic co-occurrence in the automotive public opinion data, and the node color is determined based on the Beta value of the topic keyword in the topic set.