Theme-based automobile public praise data analysis method, system and device and medium

Abstract

The invention provides a theme-based automobile public praise data analysis method, system and device and a medium, and relates to the technical field of public praise data analysis. According to the method, aggregated word-of-mouth text data related to a specific automobile theme is collected and preprocessed to generate structured text data. Secondly, converting the data into text feature vectors by adopting embedded representation learning, and constructing a topic distribution diagram containing semantic association, affiliation and time association on the basis of the text feature vectors; and finally, fusing information of the theme distribution map and the text feature vector by using a deep learning model, and outputting an emotion score for a specific automobile theme. The method can effectively capture and analyze emotional tendencies in a large amount of user generated contents, and provides more accurate and comprehensive emotional evaluation for a specific automobile theme.

Description

Technical Field

[0001] This invention relates to the field of reputation data analysis technology, and in particular to topic-based automotive reputation data analysis methods, systems, devices, and media. Background Technology

[0002] Thematic word-of-mouth data is typically characterized by high timeliness, high noise levels, and scattered themes. It may also contain complex emotions (such as sarcasm and irony) and be a mixture of multiple themes. Current analysis methods mostly process single texts, such as analyzing only one comment under a theme. This approach can easily lead to unbalanced analysis results and fail to reflect the differences in content and text representation caused by periodic changes in the data. Summary of the Invention

[0003] The purpose of this invention is to provide a method, system, apparatus, and medium for topic-based automotive reputation data analysis to solve one or more technical problems existing in the prior art, or at least provide a beneficial alternative or create conditions.

[0004] The solution to the technical problem of this invention is as follows: On the one hand, this invention provides a topic-based automotive reputation data analysis method, including the following steps: Collect aggregated word-of-mouth text data related to specific car topics; wherein, the aggregated word-of-mouth text data consists of the original topic posts with timestamps and their corresponding reply posts published on specific car topics; The aggregated word-of-mouth text data is preprocessed to obtain structured text data; Based on the structured text data, embedding representation learning is performed to obtain text feature vectors; Construct a topic distribution map based on the text feature vectors; The nodes of the topic distribution graph represent text units in the structured text data, and the edges include semantic association edges, subordinate relationship edges, and temporal association edges between the nodes. Based on the topic distribution map and the text feature vector, a deep learning model is used for fusion processing to output a sentiment score related to a specific car topic.

[0005] Furthermore, the collection of aggregated word-of-mouth text data related to specific automotive topics includes the following steps: Obtain the original topic posts under a specific car topic, wherein the original topic posts are selected from at least one of forum main posts, video comment section topic posts, and social media topic posts, and record their data sources; Retrieve the reply posts corresponding to the original topic post; the reply posts include first-level reply posts, multi-level nested reply posts, and corresponding interaction data.

[0006] Furthermore, the preprocessing of the aggregated word-of-mouth text data to obtain structured text data includes the following steps: The aggregated word-of-mouth text data is deduplicated to filter out redundant text with a duplication rate exceeding a preset duplication rate threshold. The aggregated word-of-mouth text data after deduplication is associated and stored according to the car-themed category tags to generate the structured text data; The structured text data includes the topic number, data source, text content, and interaction data corresponding to the original topic post and the reply post.

[0007] Furthermore, the text feature vector includes low-dimensional feature vectors of the topic post and low-dimensional feature vectors of the reply post; The step of learning embedding representations based on the structured text data to obtain text feature vectors includes the following steps: Based on the structured text data, the text content of the topic post and the reply post is extracted; The text content of the topic post is input into a pre-trained language model for encoding to generate an initial embedding vector for the topic post; wherein, the language model is a BERT-base model adjusted based on automotive domain corpus; The text content of the topic post is concatenated with the text content of the corresponding reply post to form the topic reply context text, which is then input into the language model for encoding to generate the initial embedding vector of the reply post. UMAP dimensionality reduction is performed on the initial embedding vector of the topic post and the initial embedding vector of the reply post to obtain the low-dimensional feature vector of the topic post and the low-dimensional feature vector of the reply post.

[0008] Furthermore, the nodes of the topic distribution map include topic post nodes and reply post nodes, and each node is associated with a corresponding text feature vector; The step of constructing a topic distribution map based on the text feature vector includes the following steps: Semantic association edges are established between node pairs whose cosine similarity to the text feature vectors is greater than a preset similarity threshold, and their weights are set to the normalized value of the cosine similarity. Establish a dependency relationship edge between the topic post node and its corresponding reply post node, and set its weight to 1; Extract the publication timestamp of the text unit, calculate the publication time difference of each node, establish a time association edge between node pairs whose publication time difference is within a preset time difference threshold, and introduce a time decay coefficient to dynamically adjust the weight of the time association edge. The topic distribution graph is constructed by integrating the semantic association edges, the subordinate relationship edges, and the temporal association edges.

[0009] Furthermore, the weights of the time-related edges satisfy the following calculation formula: ; in, Indicates the first The node and the first The weights of time-related edges between nodes; Represents a mapping function; Indicates the first The release time of each node, Indicates the first The release time of each node, time The corresponding time decay coefficient; The time decay coefficient satisfies the following calculation formula: ; in, This indicates the initial value of the preset time decay coefficient. Indicates a time period.

[0010] Furthermore, the deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network, and a third graph convolutional neural network; the third graph convolutional neural network includes multiple graph convolutional layers and uses a preset dropout rate for Dropout regularization; The step of fusing the topic distribution map and the text feature vector using a deep learning model to output a sentiment score related to a specific car topic includes the following steps: Based on the semantic association edges, subordinate relationship edges, and temporal association edges of the topic distribution graph, three adjacency relationship matrices are established and input into the deep learning model. The graph attention network is used to perform interactive enhancement embedding on the adjacency relationship matrix, and learnable attention weights are assigned to each node in the topic distribution graph. The features of neighbor nodes under different relationships are aggregated through a multi-head attention mechanism to output relationship enhancement features. The relationship enhancement features are input into the first graph convolutional neural network, and layered convolution is performed to output layered fusion features. The text feature vector and the hierarchical fusion feature are input into the second graph convolutional neural network for hybrid feature propagation, and the output is a comprehensive feature. The comprehensive features are input into the third graph convolutional neural network for deep feature propagation, and multi-layer deep graph features are output. The mean of the deep graph features in each layer is used as the sentiment score.

[0011] On the other hand, this application provides a topic-based automotive reputation data analysis system, including a data collection module, a topic distribution map module, and a sentiment score calculation module; The data collection module is used to collect aggregated word-of-mouth text data related to a specific car topic; wherein, the aggregated word-of-mouth text data is the text data of the original topic post with timestamps and its corresponding reply posts published on the specific car topic; The topic distribution graph module is used to preprocess the aggregated word-of-mouth text data to obtain structured text data; based on the structured text data, embedding representation learning is performed to obtain text feature vectors; based on the text feature vectors, a topic distribution graph is constructed; wherein, the nodes of the topic distribution graph represent text units in the structured text data, and the edges include semantic association edges, subordinate relationship edges, and time association edges between the nodes; The sentiment score calculation module is used to perform fusion processing based on the topic distribution map and the text feature vector using a deep learning model, and output a sentiment score related to a specific car topic.

[0012] On the other hand, this application provides a topic-based automotive reputation data analysis device, including a processor and a memory; the memory stores a computer program, and the processor executes the computer program to implement the aforementioned topic-based automotive reputation data analysis method.

[0013] On the other hand, this application provides a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the aforementioned topic-based automotive reputation data analysis method.

[0014] The beneficial effects of this invention are as follows: This application provides a topic-based automotive reputation data analysis method. It collects aggregated reputation text data related to specific automotive topics and preprocesses it to generate structured text data. Then, it uses embedding representation learning to convert this data into text feature vectors, and constructs a topic distribution map containing semantic associations, subordinate relationships, and temporal associations based on these vectors. Finally, it uses a deep learning model to fuse the information from the topic distribution map and the text feature vectors, outputting a sentiment score for a specific automotive topic. This method can effectively capture and analyze the sentiment tendencies in a large amount of user-generated content, providing a more accurate and comprehensive sentiment assessment for specific automotive topics. This application also provides corresponding systems, devices, and media. The beneficial effects of the systems, devices, and media are similar to those of the method and will not be elaborated here.

[0015] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0017] Figure 1 This is a flowchart of the topic-based automotive reputation data analysis method provided in this application; Figure 2 This is a structural diagram of the topic-based automotive reputation data analysis system provided in this application; Figure 3 This is a structural diagram of the topic-based automotive reputation data analysis device provided in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0020] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Thematic word-of-mouth data is typically characterized by its timeliness, high noise levels, and scattered themes, and may contain complex emotions (such as sarcasm and irony) and mix of multiple themes. This data is crucial for analyzing user attitudes towards specific topics or products, especially in sectors like the automotive industry, where user discussion preferences are closely related to specific product experiences. Therefore, effectively extracting valuable information from large amounts of user-generated content has become an important research direction.

[0023] Currently, the main methods for analyzing topic-based word-of-mouth data include single-text analysis methods, such as analyzing a single comment under a specific topic. These methods typically use natural language processing techniques to extract text features and leverage machine learning algorithms for classification or regression tasks to predict the sentiment or other attributes of the text.

[0024] However, existing analytical methods have several significant drawbacks. Traditional methods often focus only on individual text units (such as a single comment), ignoring the connections between texts and overall structural information, making it difficult to comprehensively reflect the true attitudes of user groups. Due to the lack of effective consideration of the time dimension, existing methods struggle to reflect the trends and periodic fluctuations of data over time, thus affecting the accuracy of long-term trend analysis. Existing methods perform poorly when dealing with complex emotional expressions such as sarcasm and irony, and also struggle to accurately distinguish the boundaries between different topics when faced with mixed comments involving multiple themes. In practical applications, topical word-of-mouth data often contains a large amount of noise (such as advertisements, duplicate content, special characters, etc.), which interferes with the analysis results and reduces the accuracy of the model.

[0025] In view of the above challenges, this application proposes a topic-based automotive reputation data analysis method, system, device, and medium, which optimizes feature representation through joint topic sentiment modeling. This method not only effectively filters out noise but also integrates the temporal dimension and semantic relationships between texts, thereby more accurately calculating sentiment-related automotive industry topic semantic clusters, reducing the impact of chaotic user reviews in trending topics on sentiment analysis, and enhancing the significance of user analysis results. This method significantly improves the ability to capture complex interaction structures while maintaining high computational efficiency by constructing a topic distribution map and utilizing a deep learning model to fuse different types of related information.

[0026] First, the topic-based automotive reputation data analysis method provided in this application will be described in detail below with reference to the accompanying drawings.

[0027] Reference Figure 1 The implementation process of the topic-based automotive reputation data analysis method provided in this application embodiment includes, but is not limited to, the following steps.

[0028] Step S110: Collect aggregated word-of-mouth text data related to a specific car topic.

[0029] Among them, the aggregated word-of-mouth text data consists of the original topic posts with timestamps and their corresponding reply posts published on a specific automotive topic.

[0030] In step S110, the text data of the original topic posts and their corresponding replies for a specific automotive topic are obtained. This data includes not only direct user feedback (such as opinions on a particular car model, user experience, etc.), but also responses and discussions from other users, forming a dialogue network around the specific topic. By recording these timestamped data sources, it is possible to ensure that the chronological order of information and its dynamic changes are taken into account during subsequent analysis, thereby better understanding the evolution of public attitudes towards the topic.

[0031] Step S120: Preprocess the aggregated word-of-mouth text data to obtain structured text data.

[0032] In step S120, after collecting a large amount of raw data, it needs to be cleaned and structured to facilitate subsequent analysis. This includes removing duplicate content, irrelevant characters, and non-textual information (such as images). Simultaneously, the remaining text data is categorized according to the automotive theme and stored in association. This process helps reduce noise interference, improves data quality, and ensures that each piece of information clearly corresponds to a specific theme category, providing a guarantee for building a more accurate model.

[0033] Step S130: Based on the structured text data, perform embedding representation learning to obtain text feature vectors.

[0034] In step S130, a pre-trained language model (such as a BERT-base model adjusted based on automotive domain corpora) is used to transform the structured text into numerical feature vectors. This not only converts natural language into a machine-understandable form, but also further optimizes these vectors through UMAP dimensionality reduction technology, enabling them to better capture deep semantic information in the text. This effectively extracts key information from each post and reply, providing the necessary input for the subsequent construction of the topic distribution map.

[0035] Step S140: Construct a topic distribution map based on the text feature vector.

[0036] In the topic distribution graph, the nodes represent text units in the structured text data, and the edges include semantic association edges, subordinate relationship edges, and time association edges between nodes.

[0037] In step S140, a topic distribution graph reflecting the relationships between text units is created based on the text feature vectors generated in the previous step. This graph consists of nodes and edges, where nodes represent text units (i.e., topic posts or reply posts), and edges represent the semantic, subordinate, and temporal relationships between these nodes. This graphical representation not only intuitively displays the connections between various pieces of information but also reveals the complex patterns hidden behind the data, laying the foundation for in-depth analysis of sentiment.

[0038] Step S150: Based on the topic distribution map and text feature vectors, a deep learning model is used for fusion processing to output the sentiment score related to the specific car topic.

[0039] In step S150, the topic distribution map and text feature vectors are comprehensively processed using a deep learning model (such as graph attention networks and graph convolutional neural networks). This model can automatically learn and identify different types of relationships and calculate the sentiment score for each topic accordingly. This process takes into account the semantic logic within the text and the interactions between them, thus providing more accurate and comprehensive sentiment assessment results than traditional methods. This approach not only improves the accuracy of sentiment analysis but also provides strong support for decision-makers.

[0040] In some embodiments of this application, the process of collecting aggregated word-of-mouth text data related to a specific car topic in step S110 includes, but is not limited to, the following steps.

[0041] Step S210: Obtain the original topic posts under the specific car topic. The original topic posts are selected from at least one of the forum main posts, video comment section topic posts, and social media topic posts, and record their data sources.

[0042] In step S210, original topic posts related to a specific car topic are identified and obtained from multiple sources (such as forum main posts, video comment section topic posts, and social media topic posts), and the data source for each post is recorded. This method ensures that the collected data is not only broad in coverage but also highly representative, comprehensively reflecting users' opinions and attitudes towards a specific car topic on different platforms. Simultaneously, recording the data source facilitates source verification or further data mining during subsequent analysis, improving the data's credibility and traceability.

[0043] Step S220: Retrieve the reply posts corresponding to the original topic post. Reply posts include first-level reply posts, multi-level nested reply posts, and corresponding interaction data.

[0044] In step S220, after obtaining the original topic posts, all replies under these posts are captured, including first-level replies, multi-level nested replies, and interaction data between them (such as the number of likes and dislikes). This is done to capture the complete dialogue flow around a specific topic and understand the interaction and influence between different user viewpoints. Furthermore, considering that replies may contain support, rebuttals, or supplementary explanations for the original post's viewpoint, this information is crucial for a comprehensive understanding of public opinion on the topic. By capturing multi-level reply data, a richer and more detailed topic discussion network can be constructed, providing a solid foundation for subsequent sentiment analysis and trend prediction.

[0045] In some embodiments of this application, the process of preprocessing the aggregated word-of-mouth text data to obtain structured text data in step S120 includes, but is not limited to, the following steps.

[0046] Step S310: Perform deduplication on the aggregated word-of-mouth text data and filter out redundant text with a duplication rate exceeding a preset duplication rate threshold.

[0047] In step S310, the quality of the dataset is improved by removing duplicate text information. Specifically, redundant text with a duplication rate exceeding a preset threshold is filtered out to reduce noise and unnecessary data redundancy. Deduplication not only saves storage space but also avoids biases or misleading results caused by duplicate data in subsequent analyses. This is crucial for ensuring the authenticity and reliability of the analysis results, especially when dealing with large-scale user-generated content, as this process helps improve the overall effectiveness of the dataset.

[0048] Step S320: The deduplicated aggregated word-of-mouth text data is associated and stored according to the car-themed category tags to generate structured text data.

[0049] The structured text data includes the topic number, data source, text content, and interaction data corresponding to the original topic post and reply post.

[0050] In step S320, after deduplication, the cleaned data is categorized according to its automotive theme and stored in conjunction with corresponding metadata (such as theme ID, data source, text content, and interaction data) to form a structured text dataset. This structured processing method allows each piece of information to be accurately located under a specific theme category, greatly improving the efficiency of data retrieval and analysis. Furthermore, by recording detailed metadata, researchers can delve deeper into the connections between different themes and the interaction patterns between users, providing strong data support for subsequent sentiment analysis and trend prediction. This step is a crucial step in transforming raw, messy data into ordered, analyzable data.

[0051] In some embodiments of this application, the text feature vector includes low-dimensional feature vectors of topic posts and low-dimensional feature vectors of reply posts. Step S130, which involves performing embedding representation learning based on structured text data to obtain the text feature vector, includes, but is not limited to, the following steps.

[0052] Step S410: Extract the text content of the topic post and reply post based on the structured text data.

[0053] In step S410, the original text content for semantic modeling is precisely extracted from the preprocessed structured text data. Specifically, the independent text content of the original topic post and the text content of all related reply posts need to be extracted separately. This extraction process not only preserves the semantic subject of the text but also lays the foundation for subsequent construction of context-aware semantic representations. By separating and clearly identifying the content of topic posts and reply posts, it is possible to ensure differentiated processing of text at different levels during subsequent encoding, thereby more accurately capturing the semantic hierarchy and information flow direction in user discussions and providing high-quality input for achieving fine-grained sentiment and topic analysis.

[0054] Step S420: Input the text content of the topic post into the pre-trained language model for encoding to generate the initial embedding vector of the topic post.

[0055] The language model is a BERT-base model adjusted based on automotive domain corpus.

[0056] In step S420, a deep semantic model is used to independently encode the topic posts to generate high-dimensional semantic vectors. A BERT-base model, fine-tuned based on automotive domain corpora, is employed to enhance its domain adaptability and more accurately understand automotive-related terminology, expression habits, and sentiment. After inputting the topic post content into the model, it extracts context-related semantic features through a multi-layer Transformer structure. The output initial embedding vector not only contains the semantic information of the words themselves but also incorporates the influence of the context. This step is crucial for constructing topic semantic anchors. As the starting point of the discussion, the topic post's embedding vector will act as a core node in the subsequent graph structure, directly influencing the judgment of the overall topic's sentiment and direction.

[0057] Step S430: Concatenate the text content of the topic post with the text content of the corresponding reply post to form the topic reply context text, input it into the language model for encoding, and generate the initial embedding vector of the reply post.

[0058] In step S430, a context-aware semantic representation of the reply post is constructed. Since user replies often rely on the content of the original topic post, encoding the reply text in isolation may lose its semantic context, leading to semantic ambiguity or misunderstanding. Therefore, by concatenating the topic post content and the reply post content to form a joint "topic-reply" contextual text, and then inputting it into the same language model for encoding, the model can fully consider its semantic relationship with the topic when generating the reply embedding vector. This approach effectively captures the contextual dependence of replies, enhances the understanding of complex emotional expressions such as irony, sarcasm, support, or opposition, thereby improving the accuracy and richness of the reply semantic representation.

[0059] Step S440: Perform UMAP dimensionality reduction on the initial embedding vector of the topic post and the initial embedding vector of the reply post respectively to obtain the low-dimensional feature vector of the topic post and the low-dimensional feature vector of the reply post.

[0060] In step S440, although the embedding vectors generated by the language model possess rich semantic information, their high-dimensionality (such as BERT's 768 dimensions) may lead to high computational complexity, low training efficiency, and the "curse of dimensionality" in subsequent graph structure construction and model training. Therefore, this step introduces UMAP (Uniform Manifold Approximation and Projection) dimensionality reduction technology to perform non-linear dimensionality reduction on the initial embedding vectors. UMAP significantly reduces the vector dimensionality while preserving the local and global structure in the high-dimensional space, ensuring that semantically similar texts maintain a similar distribution in the low-dimensional space. This not only improves computational efficiency but also enhances the interpretability and clustering performance of the vectors, providing a compact and semantically clear low-dimensional feature foundation for subsequently constructing node representations in the topic distribution graph.

[0061] In some embodiments of this application, the first One thread The text content is input into the pre-trained language model BERT for encoding, generating initial embedding vectors for topic posts. ,Right now For the first One thread Corresponding reply post , will the topic post Text content and replies The text content of the reply posts is concatenated to form the context text of the reply to the topic. The input is fed into the BERT language model for encoding, generating the initial embedding vector of the reply post. ,Right now ; Initial embedding vector for topic posts and the initial embedding vector of the reply post Perform UMAP dimensionality reduction processing to obtain the low-dimensional feature vectors of the topic posts. And the low-dimensional feature vectors of reply posts .

[0062] In some embodiments of this application, the topic posts are assigned according to the weighting rules. Text content and replies The text content of the reply posts is concatenated. The weight allocation rules include time-series dynamic weight allocation rules based on LSTM networks, context-based dynamic weight allocation rules based on self-attention mechanisms, and fixed weight allocation rules based on expert experience.

[0063] Specifically, the temporal dynamic weight allocation rule based on LSTM networks utilizes LSTM networks to capture the temporal sequence characteristics between topic posts and replies, thereby dynamically assigning weights to each text feature vector. By incorporating temporal order, this method can identify importance patterns that change over time. For example, replies posted early in a discussion may be more significant for understanding the overall trend of the topic, while later replies may be more supportive or defiance of existing viewpoints. The LSTM-based model can learn these time-varying differences in importance and adjust the weights accordingly, making the final generated text feature vectors more reflective of the true dynamic process of the discussion.

[0064] The context-dynamic weighting rule based on self-attention employs a self-attention mechanism to evaluate and assign weights, focusing on the contextual relationships between text content. Self-attention allows the model to automatically adjust its attention based on the relevance of different parts of the input sequence, thus assigning higher weights to content more relevant to the current discussion topic. This approach is particularly suitable for handling situations involving complex sentiment expressions or mixed topics, as it can identify and emphasize details crucial for understanding the overall sentiment. In this way, not only is the quality of text representation improved, but subtle differences in user discussions are also captured more accurately, enhancing the accuracy of sentiment analysis.

[0065] In a fixed-weight allocation rule based on expert experience, weights are pre-set fixed values ​​based on the experience of experts in the domain. This method relies on experts' understanding and judgment of a specific automotive topic and related discussions to determine which types of posts or replies should be given higher importance. For example, if an expert believes that the original topic post is generally more informative than the replies, they might assign a higher fixed weight to the topic post. While this method lacks flexibility and cannot adapt to the characteristics of the data itself like the previous two rules, it provides a simple and direct way to allocate weights, especially suitable for small datasets or when rapid model deployment is required without sufficient training data. Furthermore, fixed-weight allocation rules can serve as a benchmark to compare the effectiveness of other, more complex dynamic weight allocation strategies.

[0066] In some embodiments of this application, the nodes of the topic distribution graph include topic post nodes and reply post nodes, and each node is associated with a corresponding text feature vector. Step S140, the process of constructing the topic distribution graph based on the text feature vector, includes, but is not limited to, the following steps.

[0067] Step S510: Establish semantic association edges between node pairs whose cosine similarity to the text feature vector is greater than a preset similarity threshold, and set their weights to the normalized value of the cosine similarity.

[0068] In step S510, the semantic proximity of nodes (i.e., topic posts and reply posts) is measured by calculating the cosine similarity. When the cosine similarity between the text feature vectors of two nodes exceeds a set threshold (e.g., 0.7), a semantic connection edge is established between the two nodes, and its weight is set to the normalized value of that similarity. This connection method based on semantic similarity helps to capture the inherent connections between different posts in the discussion, especially those posts that are highly semantically related but have little direct interaction. The network constructed in this way can reveal the deep structure hidden in the data, which is crucial for understanding the overall sentiment and propagation path of a topic.

[0069] Step S520: Establish a subordinate relationship edge between the topic post node and its corresponding reply post node, and set its weight to 1.

[0070] Step S520 emphasizes the direct relationship between the main thread and its replies, meaning each reply is a response to a specific main thread. To reflect this hierarchical relationship, dependency edges are established between the main thread node and all its reply thread nodes, with the weight of these edges uniformly set to 1. This approach not only clarifies the hierarchical structure of the discussion tree but also provides a clear relational framework for subsequent analysis. The existence of dependency edges makes it easier for the model to track the direction of information flow, understand how each reply revolves around the original topic, and how the entire discussion gradually develops.

[0071] Step S530: Extract the publication timestamp of the text unit, calculate the publication time difference of each node, establish time association edges between node pairs whose publication time difference is within the preset time difference threshold, and introduce a time decay coefficient to dynamically adjust the weight of the time association edges.

[0072] In step S530, considering the impact of time on the dynamics of the discussion, the temporal distance between nodes (posts) is assessed by comparing their publication timestamps. If the publication time difference between two nodes is less than a preset time difference threshold (e.g., 36 hours), they are considered related in the time dimension, and a time-related edge is established between the two nodes. Furthermore, to reflect the time decay effect—that is, newer replies may better reflect the current sentiment state than older replies—a time decay coefficient is introduced to dynamically adjust the weights of the time-related edge. This method allows the model to more accurately capture the changing trends in discussion popularity and the evolution of public sentiment over time.

[0073] Step S540: Integrate semantic association edges, subordinate relationship edges, and temporal association edges to construct a topic distribution graph.

[0074] In step S540, all the edges of the above types are integrated to form a comprehensive topic distribution graph. This graph includes not only the semantic and hierarchical relationships between posts, but also their temporal relationships, thus providing a multi-dimensional perspective on the discussion network. The topic distribution graph constructed in this way can more accurately reflect the true state of user discussions, better showcasing both sentiment trends and topic diffusion patterns. This provides a solid foundation for subsequent advanced analyses such as sentiment score prediction using deep learning models, and also improves the ability to understand complex interaction structures.

[0075] In some embodiments of this application, the weights of time-related edges satisfy the following calculation formula (1): (1); In formula (1), Indicates the first The node and the first The weights of time-related edges between nodes. This represents a mapping function. Indicates the first The release time of each node, Indicates the first The release time of each node, time The corresponding time decay coefficient.

[0076] In some embodiments of this application, the time decay coefficient satisfies the following calculation formula (2): (2); In formula (2), This indicates the initial value of the preset time decay coefficient (e.g., 0.5). Indicates a time period.

[0077] In some embodiments of this application, the deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network, and a third graph convolutional neural network. The third graph convolutional neural network includes multiple graph convolutional layers and performs Dropout regularization using a preset dropout rate. In step S150, the process of fusing the topic distribution map and text feature vectors using the deep learning model to output a sentiment score related to a specific car topic includes, but is not limited to, the following steps.

[0078] Step S610: Based on the semantic association edges, subordinate relationship edges, and temporal association edges of the topic distribution graph, establish three adjacency relationship matrices and input them into the deep learning model.

[0079] Step S610 is a crucial step in formalizing the multidimensional relationships in the constructed topic distribution graph. Specifically, adjacency matrices are established for the three different types of edges in the graph—semantic association edges, subordinate relationship edges, and temporal association edges. Each adjacency matrix precisely characterizes the connection pattern between nodes under a specific relationship: the semantic association matrix reflects the similarity of text content in the semantic space; the subordinate relationship matrix reflects the hierarchical structure between topic posts and reply posts; and the temporal association matrix captures the dynamic evolution of user discussions over time. These three independent adjacency matrices are fed as parallel inputs into the subsequent deep learning model, providing the model with structured, multi-perspective relational information. This enables the model to distinguish and fuse different types of relational signals, thereby gaining a more comprehensive understanding of the information dissemination and emotional interaction mechanisms in complex word-of-mouth networks.

[0080] Step S620: A graph attention network is used to perform interactive enhancement embedding on the adjacency relationship matrix. Learnable attention weights are assigned to each node in the topic distribution graph. The features of neighbor nodes under different relationships are aggregated through a multi-head attention mechanism to output relationship enhancement features.

[0081] In step S620, a Graph Attention Network (GAT) is introduced to dynamically weight and enhance the features of the multi-relationship graph structure. Traditional graph convolution methods typically treat all neighbor nodes equally or use fixed weights. This step, however, uses an attention mechanism to assign learnable attention weights to each node in the topic distribution graph, enabling the model to automatically determine which relationship (semantic, subordinate, or temporal) is more important in a specific context. For example, during periods of heightened emotion, temporal associations may be more predictive; while in deep discussions, semantic associations are more crucial. The multi-head attention mechanism further enhances the model's expressive power by computing multiple attention heads in parallel, capturing relationship patterns in different subspaces, and aggregating the results to generate more discriminative "relationship-enhancing features." This step represents a leap from static graph structures to dynamic semantically aware graph representations, significantly enhancing the model's ability to model complex interaction patterns.

[0082] Step S630: Input the relationship enhancement features into the first graph convolutional neural network, perform hierarchical convolution, and output the hierarchical fusion features.

[0083] In step S630, after obtaining the multi-relationship features enhanced by the attention mechanism, a first graph convolutional neural network (GCN) is used to perform hierarchical aggregation and information propagation. Graph convolution operations integrate local information by weighted summation of the features of each node and the features of its neighboring nodes through an adjacency matrix. Through iterative propagation across multiple layers of GCN, node features gradually incorporate information from multi-hop neighbors, thereby capturing a broader range of contextual dependencies in the graph. The "hierarchical fusion features" output by this process not only contain the semantic information of the node itself but also incorporate higher-order neighbor information in semantic, structural, and temporal dimensions, forming a deep representation of the node's role and influence in the overall discussion network. This layer of convolution focuses on feature extraction within the relational structure, laying the foundation for subsequent cross-modal fusion.

[0084] Step S640: Input the text feature vector and the hierarchical fusion feature into the second graph convolutional neural network, perform hybrid feature propagation, and output the comprehensive feature.

[0085] In step S640, deep fusion of original text semantic information and graph structure evolution information is achieved. In the previous steps, the text feature vectors retain the original semantics without graph propagation, while the hierarchically fused features have undergone graph attention and the first round of GCN processing, enriching them with structural context information. By feeding both of these as inputs into the second graph convolutional neural network, the model performs a "hybrid feature propagation" operation, that is, propagating the original text features and the fused structural features simultaneously on the graph structure. This design avoids early information loss, ensures the fidelity of the original semantics, and allows the original features to undergo secondary propagation and optimization within the graph structure. The introduction of techniques such as LayerNorm further stabilizes the training process. The final output "comprehensive features" are an organic unity of original semantics and high-order structural information, possessing stronger robustness and expressive power, providing high-quality input for the final sentiment judgment.

[0086] Step S650: Input the comprehensive features into the third graph convolutional neural network, perform deep feature propagation, output multi-layer deep graph features, and use the average value of each layer of deep graph features as the sentiment score.

[0087] In step S650, the third graph convolutional neural network is composed of multiple stacked GCNs and employs Dropout regularization to prevent overfitting and enhance the model's generalization ability. This network performs deep propagation of the "comprehensive features," extracting higher-level abstract features layer by layer. The output of each GCN layer represents a graph representation under different receptive fields: shallow layers capture local neighborhood information, while deeper layers reflect the global network structure.

[0088] Unlike traditional approaches that only use the output of the last layer, this step innovatively averages the features of the depth map from multiple layers as the final sentiment score. This "jump connection" averaging strategy (similar to DenseNet or Jumping Knowledge) fuses representations from different layers, balancing local details with global semantics, effectively mitigating the information smoothing problem in deep networks. Finally, this average vector, after being mapped by an activation function (such as Sigmoid or Tanh), outputs a continuous sentiment score for a specific car topic, comprehensively, stably, and interpretably reflecting the sentiment tendency of public opinion.

[0089] In some embodiments of this application, firstly, three adjacency matrices are constructed based on the semantic association edges (s), subordinate relationship edges (c), and temporal association edges (t) of the topic distribution graph, wherein, This represents the adjacency matrix established based on semantically related edges (s). This represents the adjacency matrix established based on the subordinate edge (c). This represents the adjacency matrix established based on time-related edges (t).

[0090] for Its element is defined as: if and only if the node and There are edges between them. ,in express and The weights corresponding to the types of edges between them. This is obtained by adding self-joins to the adjacency matrix. And calculate the degree matrix. .

[0091] Secondly, a graph attention network is used to compute nodes based on three types of relation edges. and Learnable attention weights between It satisfies the following calculation formula (3): (3); In formula (3), Represents a node Basic Embedded Features enhanced by graph attention network, among which This represents a graph attention network. , Represents a node The corresponding text unit; similarly, we can obtain, Represents a node Basic Embedded Features enhanced by graph attention network Represents a node Basic Embedded Features enhanced by graph attention networks; Represented as a learnable parameter matrix, Represents the attention coefficient vector. Represents a node The neighbor set contains three relation types, namely .

[0092] Furthermore, a multi-head attention mechanism is used to aggregate neighbor features and output the node. In the Layer Relationship Enhancement Features It satisfies the following calculation formula (4): (4); In formula (4), Represents a node neighboring nodes In the Layer relationship enhancement features, Indicates the first For relation types in layer graph convolution The learnable parameter matrix, This represents the ReLU activation function.

[0093] The first convolutional neural network consists of multiple convolutional layers. Relationship enhancement features are input into the first convolutional neural network, undergoing hierarchical convolution, and the output is the first... Layered fusion features As It satisfies the following formula (5): (5); In formula (5), Represents the normalized adjacency matrix. Indicates the first Layered fusion features output from convolutional layers The initial hierarchical fusion features are composed of relationship enhancement features of each node. Indicates the first The learnable parameter matrix of each convolutional layer.

[0094] Then, the text feature vector and With layered fusion features The input is fed into the second convolutional neural network, where mixed features are propagated, and the output is a comprehensive feature. It satisfies the following calculation formula (6): (6); in, It consists of the basic embeddings of all nodes. The structured embedding is represented by low-dimensional feature vectors of topic posts and low-dimensional feature vectors of reply posts; Representation layer normalization, This indicates a feature splicing operation.

[0095] Then, the comprehensive features The input is fed into the third convolutional neural network for deep feature propagation, and the output is the depth map features of each layer, which satisfy the following calculation formula (7): (7); In formula (7), Indicates the first Depth map features of the layer Indicates the first The depth map features of layer 1 are as follows: As input.

[0096] Finally, the mean value of the features from each depth map layer is used as the sentiment score. It satisfies the following calculation formula (8): (8); In formula (8), This indicates the number of layers in the convolutional neural network in the third graph, typically set to 2 or 3 layers.

[0097] In summary, the topic-based automotive reputation data analysis method provided in this application has the following technical effects.

[0098] This method first involves comprehensive data collection and preprocessing, acquiring original topic posts and their replies from multiple sources (such as forums, video comment sections, and social media topic posts), and performing deduplication and structuring to ensure data quality and representativeness. A pre-trained language model is used to encode the text content, and these vectors are optimized using UMAP dimensionality reduction technology to achieve efficient embedding representation learning, thereby capturing deep semantic information in the text. Next, by constructing a topic distribution graph containing semantic association edges, subordinate relationship edges, and temporal association edges, this method can effectively model the complex relationships between different posts in the discussion, revealing not only the information dissemination path but also reflecting the changing trends of user attitudes over time.

[0099] Furthermore, by combining graph attention networks and multi-layer graph convolutional neural networks, this method achieves dynamic weighting and deep fusion of various types of relationship features, improving the accuracy of sentiment analysis. Compared to traditional single-text analysis methods, this method more comprehensively considers the relationships between texts, time factors, and the influence of the overall discussion environment, providing more accurate and reliable sentiment analysis results. In addition, by averaging graph features at different levels as the final output, the interpretability and robustness of the model are enhanced, providing a powerful data-driven tool for decision support in the automotive industry and related fields.

[0100] Secondly, refer to Figure 2 This application provides a topic-based automotive reputation data analysis system, including a data collection module 710, a topic distribution map module 720, and an emotion score calculation module 730.

[0101] The data collection module 710 is used to collect aggregated word-of-mouth text data related to a specific automotive topic. This aggregated word-of-mouth text data consists of the original topic posts with timestamps and their corresponding replies published on the specific automotive topic.

[0102] The topic distribution graph module 720 is used to preprocess the aggregated word-of-mouth text data to obtain structured text data. Based on the structured text data, embedding representation learning is performed to obtain text feature vectors. A topic distribution graph is then constructed based on these text feature vectors. In this graph, nodes represent text units within the structured text data, and edges include semantic association edges, dependency edges, and temporal association edges between nodes.

[0103] The sentiment score calculation module 730 is used to perform fusion processing based on the topic distribution map and text feature vector using a deep learning model, and output the sentiment score related to a specific car topic.

[0104] Furthermore, refer to Figure 3 This application provides a topic-based automotive reputation data analysis apparatus, including a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned automotive reputation data analysis method.

[0105] Furthermore, embodiments of this application provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the aforementioned automotive reputation data analysis method.

[0106] Similarly, the technical effects of the systems, devices, and media provided in this application embodiment are the same as those of the above method embodiments, and will not be repeated here.

[0107] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards of the relevant countries and regions. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data for the normal operation of the embodiments of this application obtained.

[0108] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0109] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0110] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0112] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or, if necessary, processing in a suitable manner, and then stored in computer memory.

[0113] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0114] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0115] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0116] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for analyzing car reputation data based on a theme, characterized in that, The method comprises the following steps: collecting aggregated reputation text data related to a specific automobile theme; wherein the aggregated reputation text data is the text data of original topic posts and corresponding reply posts published for a specific automobile theme with timestamps; preprocessing the aggregated reputation text data to obtain structured text data; performing embedding representation learning according to the structured text data to obtain a text feature vector; constructing a topic distribution graph according to the text feature vector; wherein the nodes of the topic distribution graph represent text units in the structured text data, and the edges include semantic association edges, subordinate relationship edges and time association edges between the nodes; performing fusion processing using a deep learning model according to the topic distribution graph and the text feature vector to output a sentiment score related to the specific automobile theme. 2.The theme-based automobile reputation data analysis method according to claim 1, characterized in that, The method of collecting aggregated reputation text data related to a specific automobile theme comprises the following steps: obtaining original topic posts under a specific automobile theme, wherein the original topic posts are selected from at least one of forum main posts, video comment section topic posts and social media topic posts, and the data sources thereof are recorded; grabbing reply posts corresponding to the original topic posts; the reply posts include first-level reply posts, multi-level nested reply posts and corresponding interaction data. 3.The theme-based automobile reputation data analysis method according to claim 1, characterized in that, The method of preprocessing the aggregated reputation text data to obtain structured text data comprises the following steps: performing deduplication processing on the aggregated reputation text data to filter redundant text with a repetition rate exceeding a preset repetition rate threshold; associatively storing the deduplicated aggregated reputation text data according to automobile theme classification labels to generate the structured text data; wherein the structured text data includes the topic numbers, data sources, text contents and interaction data of the original topic posts and the reply posts. 4.The theme-based automobile reputation data analysis method according to claim 1, characterized in that, The text feature vector includes a topic post low-dimensional feature vector and a reply post low-dimensional feature vector. The method of performing embedding representation learning according to the structured text data to obtain a text feature vector comprises the following steps: extracting the text contents of the topic posts and the reply posts from the structured text data; inputting the text contents of the topic posts into a pre-trained language model for encoding to generate topic post initial embedding vectors; wherein the language model is a BERT-base model adjusted based on automobile field corpus; splicing the text contents of the topic posts and the text contents of the corresponding reply posts to form topic-reply context text, and inputting the topic-reply context text into the language model for encoding to generate reply post initial embedding vectors; performing UMAP dimension reduction processing on the topic post initial embedding vectors and the reply post initial embedding vectors respectively to obtain the topic post low-dimensional feature vector and the reply post low-dimensional feature vector. 5.The theme-based automobile word-of-mouth data analysis method according to claim 1, characterized in that, The nodes of the topic distribution graph include topic post nodes and reply post nodes, and each node is associated with a corresponding text feature vector. The method of constructing a topic distribution graph according to the text feature vector comprises the following steps: establishing a semantic association edge between a pair of nodes with a cosine similarity of the text feature vectors greater than a preset similarity threshold, and setting the weight of the semantic association edge as a normalized value of the cosine similarity; establishing a subordination edge between the topic post node and the corresponding reply post node thereof, and setting a weight of the subordination edge to 1; extracting a publishing time stamp of each text unit, calculating a publishing time difference of each node, establishing a time correlation edge between a pair of nodes whose publishing time difference is within a preset time difference threshold, and introducing a time decay coefficient to dynamically adjust a weight of the time correlation edge; integrating the semantic correlation edge, the subordination edge and the time correlation edge to construct the topic distribution graph. 6.The theme-based automobile word-of-mouth data analysis method according to claim 5, characterized in that, The weight of the time correlation edge satisfies the following calculation formula: ； wherein, denotes the weight of the temporal association edge between the th node and the th node; denotes a mapping function; denotes the publication time of the th node, denotes the publication time of the th node, moment corresponding time decay coefficient; The time decay coefficient satisfies the following calculation formula: ； wherein denotes a preset starting value of the time decay coefficient, denotes a time period. 7.The theme-based automobile reputation data analysis method according to claim 1, characterized in that, The deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network and a third graph convolutional neural network; the third graph convolutional neural network includes multiple layers of graph convolutional layers and adopts a preset dropout rate for Dropout regularization; The deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network and a third graph convolutional neural network; the third graph convolutional neural network includes multiple layers of graph convolutional layers and adopts a preset dropout rate for Dropout regularization; The deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network and a third graph convolutional neural network; the third graph convolutional neural network includes multiple layers of graph convolutional layers and adopts a preset dropout rate for Dropout regularization; The deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network and a third graph convolutional neural network; the third graph convolutional neural network includes multiple layers of graph convolutional layers and adopts a preset dropout rate for Dropout regularization; The deep learning model includes a graph attention network, a first graph convolutional neural network, a second graph convolutional neural network and a third graph convolutional neural network; the third graph convolutional neural network includes multiple layers of graph convolutional layers and adopts a preset dropout rate for Dropout regularization. The data collection module is configured to collect aggregated review text data related to a specific automobile theme; the aggregated review text data is text data of original topic posts and corresponding reply posts published with respect to the specific automobile theme and having time stamps; The topic distribution graph module is configured to preprocess the aggregated review text data to obtain structured text data, perform embedding representation learning based on the structured text data to obtain a text feature vector, and construct a topic distribution graph based on the text feature vector; the nodes of the topic distribution graph represent text units in the structured text data, and the edges include semantic correlation edges, subordination edges and time correlation edges between the nodes; 8. A theme-based car reputation data analysis system, characterized by, The sentiment score calculation module is configured to utilize a deep learning model to perform fusion processing based on the topic distribution graph and the text feature vector, and output a sentiment score related to the specific automobile theme. ​ ​ ​ 9. A theme-based automobile reputation data analysis apparatus, characterized by comprising: An apparatus comprising a processor and a memory; the memory stores a computer program, and the processor executes the computer program to implement the method for analyzing automobile reputation data based on topics according to any one of claims 1 to 7.

10. A computer readable storage medium having stored therein a program which is executable by a processor, characterized in that, The program executable by the processor when executed by the processor is used to implement the method for analyzing automobile reputation data based on topics according to any one of claims 1 to 7.

Images