Text processing system based on topology analysis

Abstract

The application belongs to the field of text processing, and particularly relates to a text processing system based on topology analysis. The system collects network texts in real time through a classification collection module, combines a target discrimination model and a topic type library classification to form a real-time target text space. A text extraction module extracts topic key information and user basic information from the text space by using a deep model. An active evaluation module combines a hierarchical sentiment model to output a sentiment type, intensity, target severity score and attention change rate. A target topology graph module generates a hierarchical target distribution graph with a dynamic adjustment expansion range and color level by means of a map layer algorithm, a graph neural network and an HSV color space. A strategy module generates a solution strategy according to the distribution graph and a strategy library, and monitors and adjusts in real time until the target severity score is lower than a threshold value, so that full-process target analysis and disposal are realized.

Description

Technical Field

[0001] This invention belongs to the field of text processing, and particularly relates to a text processing system based on topology analysis. Background Technology

[0002] In today's information-saturated and rapidly evolving internet environment, online public opinion exhibits a complex pattern characterized by sudden outbreaks, diverse dissemination channels, multifaceted information, and concealed motives. Accurate sentiment analysis of online public opinion has become a crucial means of understanding public attitudes and preventing public opinion risks. Existing technologies utilize various models, such as the BiMPM model based on ELMo word vectors and Word2vec combined with CNN and LSTM, for public opinion analysis. However, these methods have significant limitations in handling complex semantics and special linguistic phenomena, making it difficult to accurately grasp the sentiment trends of public opinion. Furthermore, existing public opinion analyses often focus on the public opinion itself, neglecting the level of attention and influence on different groups. The level of public opinion attention and influence varies significantly among different age groups, genders, and regions, yet there is a lack of corresponding intuitive analysis tools such as heatmaps, hindering targeted and refined public opinion analysis. Therefore, this application proposes a text processing system based on topological analysis. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a text processing system and framework based on topology analysis. This system uses a classification and acquisition module to collect online text in real time, combining it with a target discrimination model and a topic type library to form a real-time target text space. A text extraction module uses a deep learning model to extract key topic information and basic user information. An active evaluation module, combined with a hierarchical sentiment model, outputs sentiment type, intensity, target severity score, and attention change rate. A target topology map module uses map layer algorithms, graph neural networks, and the HSV color space to generate a hierarchical target distribution map with dynamically adjustable expansion range and color levels. A strategy module generates solutions based on the distribution map and a strategy library, monitoring and adjusting in real time until the target severity score falls below a threshold, thus achieving full-process target analysis and handling.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A text processing system based on topology analysis includes: a classification and acquisition module, a text extraction module, an active evaluation module, a target topology graph module, and a strategy module;

[0006] The classification and acquisition module is used to collect online text information in real time and combine it with a preset target discrimination model and topic type library to perform target discrimination and classification, thereby obtaining a real-time target classification topic text space.

[0007] The text extraction module, based on the real-time target classification topic text space combined with the preset deep and short text extraction model, obtains the key information space of the topic and the basic information space of the collected users.

[0008] The proactive assessment module is used to obtain the sentiment type, negative sentiment intensity, target severity score, and topic attention change rate of the corresponding target at each level by combining the key information space of the topic and the basic information space of the collected users with the hierarchical sentiment assessment model.

[0009] The target topology map module is used to obtain a dynamic hierarchical target distribution map based on the topic key information space and the basic information space of the collected users, combined with the corresponding sentiment type, negative sentiment intensity, target severity score change rate and topic attention change rate at each level, through a map layer algorithm combined with graph neural network and HSV color space.

[0010] Specifically, the key information space of the topic includes the target topic type, target text keywords, relational attributes, target initial timestamp, dissemination time length, target dissemination range distribution, and target dissemination rate; the basic information space of the collected users includes the geographical location, age, gender, historical target types of attention, frequency of attention, corresponding sentiment type, and historical negative sentiment intensity of each type of target.

[0011] The classification and acquisition module includes a data acquisition unit and a target discrimination unit;

[0012] The data acquisition unit is used to collect information data from different platforms in real time, and to use the IKAnalyzer word segmenter to perform text segmentation, text tagging and stop word removal on the collected information data to obtain a preprocessed initial collection text set.

[0013] The target discrimination unit is used to judge targets based on the pre-processed initial collected text set, combined with a target discrimination model constructed by a pre-trained word vector model, random forest algorithm and preset dynamic discrimination threshold, to obtain the real-time target topic text space. The preset dynamic discrimination threshold is constructed by combining historical target information with the target severity score change rate, topic attention change rate, target propagation rate and propagation range per unit time through an association algorithm.

[0014] Specifically, the classification and acquisition module also includes target classification units;

[0015] The target classification unit is used to obtain the real-time target classification topic text space based on the real-time target topic text space, the text vector matrix obtained by the pre-trained word vector model, and the topic type library, and to label the type targets in real time through a matching algorithm. The topic type library is constructed by combining the historical three-level target text type text space obtained by the hierarchical analysis algorithm with a distributed database and a graph algorithm.

[0016] The text extraction module includes a deep topic information extraction unit and a user information extraction unit;

[0017] The deep topic information extraction unit is used to extract target text keywords, relational attributes, propagation time length, corresponding target propagation range distribution, and target propagation rate obtained based on the target propagation range distribution and propagation time length in real time, according to the real-time target topic text space corresponding to each type of target and the first entity relation extraction model.

[0018] Specifically, the text extraction module also includes a user information extraction unit; the proactive evaluation module includes a sentiment evaluation unit.

[0019] The user information extraction unit is used to extract the geographic location, age, gender, historical target types, frequency of attention, and comment text feature space of each user for each target type in real time based on the real-time target topic text space corresponding to each target type and the second entity relation extraction model.

[0020] The sentiment assessment unit is used to obtain the sentiment type of each user for each type of target based on the target text keywords, relational attributes, historical attention type targets, and the feature space of each user's comment text for each type of target, combined with the sentiment classification sub-model and sentiment type library in the pre-trained sentiment assessment model. Simultaneously, based on the assessment sub-model in the pre-trained sentiment assessment model, it obtains the negative sentiment intensity of each user for each type of target based on the attention frequency, propagation time length, corresponding target propagation range distribution, target propagation rate, and the keyword information and frequency of occurrence of sentiment-expressing information in each user's comment text feature space for each type of target. Finally, based on the attention sub-layer in the pre-trained sentiment assessment model and the negative sentiment intensity and type of each user for each type of target within a preset continuous time period, it obtains the negative sentiment intensity change sequence of each user for each type of target.

[0021] Specifically, the proactive assessment module also includes a public attention assessment unit and a comprehensive assessment unit;

[0022] The attention evaluation unit is used to obtain the attention level of each user at each geographical location to each type of target at each time by combining the frequency of attention, the length of the propagation time, the distribution of the corresponding target propagation range and the target propagation rate of each user for each type of target with a fuzzy algorithm. It also obtains the continuous change rate sequence of the attention level of each user at each geographical location to each type of target within a preset continuous time length by the change rate of attention frequency, the distribution of the corresponding target propagation range and the target propagation rate.

[0023] The comprehensive evaluation unit is used to obtain the target severity score for each type of target at each location at each time point by combining the negative sentiment intensity of all users at each location at each time point with the attention level of all users at each geographic location at each time point for each type of target. It also obtains the target severity score change rate for each type of target at each location at a preset continuous time length and the target severity score change rate for the same type of target at different locations at the same time point by based on the negative sentiment intensity change sequence and the continuous change rate sequence of attention level of all users at each geographic location for each type of target at a preset continuous time length.

[0024] Specifically, the proactive assessment module also includes a hierarchical correlation analysis unit; the target topology graph module includes a target graph construction unit;

[0025] The hierarchical association analysis unit is used to perform hierarchical association analysis based on the target severity score of each type of target at each location at each time point and the rate of change of the target severity score of each type of target at each location over a preset continuous time period, combined with the hierarchical rules constructed according to the user's geographic location, age, gender, and target type of interest, to obtain the hierarchical target status association space; the hierarchical relationship corresponding to the hierarchical rules is: user's geographic location level > gender level > age segment level > target type of interest level;

[0026] The target graph construction unit defines the initial position point of each type of target as the starting node, and each position point in the target propagation range distribution as the propagation branch node. At the same time point, the propagation connection relationship is constructed by multiplying the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point with the product of the rate of change of the target severity score of the same type of target at different positions at the same time point. Based on the starting node, propagation branch nodes and propagation connection relationship, combined with the graph algorithm, the initial target propagation node graph is constructed.

[0027] Specifically, the target topology graph module also includes a target graph mapping unit and a dynamic adjustment unit.

[0028] The target graph mapping unit is used to construct a hierarchical target severity score color map containing a color mapping function based on the target severity score at each level in the hierarchical target state association space and the HSV space using a support vector machine. Simultaneously, it constructs a propagation color map by multiplying the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point with the product of the rate of change of the target severity score corresponding to the same type of target at different locations at the same time point and the HSV space. This hierarchical target severity score color map is then mapped to each node in the initial target propagation node graph, and simultaneously mapped to the corresponding propagation connection relationship, thus constructing a hierarchical target distribution map. The hierarchical target distribution map includes the target severity scores corresponding to the user's geographic location level, gender level, age segment level, and interest type target level, and is distributed from top to bottom according to the level size.

[0029] Specifically, the target topology module also includes a dynamic adjustment unit;

[0030] The dynamic adjustment unit is used to collect in real time the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point, the product of the rate of change of the target severity score of the same type of target at different locations at the same time point, and the location distribution of the real-time propagation range. Combining the propagation path algorithm and the spatial rectangle expansion algorithm, it adjusts the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationship between different nodes in real time laterally. At the same time, it uses the rate of change of the target severity score of each type of target at each location under a preset continuous time length to adjust the target severity score and the corresponding color mapping depth of each node corresponding to each level in the hierarchical target distribution map vertically, so as to obtain a dynamic hierarchical target distribution map.

[0031] Specifically, the process of adjusting the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationships between different nodes in real time includes:

[0032] Based on the real-time collected target propagation rate and target severity score change rate, the data is normalized to the [0,1] interval using a normalization algorithm, and then the product of the two is calculated to obtain the weighted propagation connectivity relationship.

[0033] Based on the initial propagation node graph and historical data, the GraphSAGE graph neural network combined with the PageRank algorithm is used to predict potential propagation nodes and generate Top-K propagation paths;

[0034] Based on the real-time propagation range coordinates and potential propagation nodes, the range is initialized using the minimum enclosing rectangle algorithm. Then, the spatial decay weight is calculated by combining the Gaussian kernel function with the rate of change of target severity scores between different nodes. Combined with the expansion coefficient constructed by the target propagation rate per unit time, the dynamically expanded spatial rectangle range is obtained.

[0035] Specifically, the process of adjusting the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationships between different nodes in real time also includes:

[0036] Based on the edge weights corresponding to the propagation connection relationship and the HSV color space, the weights are converted into saturation and brightness through a linear mapping algorithm, and then the color depth of the propagation connection relationship is obtained through an HSV to RGB conversion algorithm.

[0037] Based on the adjusted rectangular range and color parameters, tiles are loaded via the map API and nodes and edges are rendered using WebGL hardware acceleration.

[0038] Based on the latest data and user interactions, the expansion coefficient and the edge weights corresponding to the propagation connection relationship are dynamically adjusted through reinforcement learning algorithms to obtain a horizontally updated hierarchical target distribution map.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] This invention addresses the shortcomings of existing technologies by achieving efficient target analysis through the collaborative efforts of various modules: the classification and acquisition module combines a preset model with a type library for real-time classification, ensuring the real-time nature and accuracy of target monitoring; the text extraction module uses a deep learning model to extract multi-dimensional key information about topics and user profiles; the proactive assessment module utilizes a hierarchical sentiment model to output sentiment and target severity scores at different levels, achieving refined assessment; the target topology map module uses graph neural networks and HSV color space mapping to dynamically expand the distribution range based on the target propagation rate and updates the color level in real time as the target severity score changes, forming an intuitive spatiotemporal evolution pattern; the strategy module uses a deep strategy library to generate disposal plans based on the dynamic target map, and dynamically iterates strategies until the risk threshold is below the target propagation rate, range, and target severity score, significantly improving target response efficiency and disposal accuracy, and achieving intelligent processing from real-time monitoring to dynamic resolution. Attached Figure Description

[0041] Figure 1 This is a block diagram of the text processing system based on topology analysis according to Embodiment 1 of the present invention;

[0042] Figure 2 This is a unit flow diagram of the text processing system based on topology analysis according to an embodiment of the present invention. Detailed Implementation

[0043] Example 1

[0044] Please see Figure 1 The present invention provides an embodiment of a text processing system based on topology analysis, applied to public opinion text analysis, comprising: a classification and acquisition module, a text extraction module, an active evaluation module, a target topology map module, and a strategy module;

[0045] The classification and collection module is used to collect online text information in real time and combine it with a preset target discrimination model and topic type library to perform target discrimination and classification, thereby obtaining a real-time target classification topic text space; it should be noted that the targets in this embodiment all represent public opinion; this will not be emphasized again below.

[0046] The text extraction module, based on the real-time target classification topic text space combined with a pre-set deep and short text extraction model, obtains the topic key information space and the basic information space of the collected users. The topic key information space includes the target topic type, target text keywords, relational attributes, target initial timestamp, propagation time length, target propagation range distribution, and target propagation rate. The basic information space of the collected users includes the geographical location, age, gender, historical target types, frequency, corresponding sentiment type, and historical negative sentiment intensity for each type of target.

[0047] The proactive assessment module is used to obtain the sentiment type, negative sentiment intensity, target severity score, and topic attention change rate of the corresponding target at each level by combining the key information space of the topic and the basic information space of the collected users with a hierarchical sentiment assessment model. Furthermore, in this embodiment, the negative sentiment intensity is marked according to the labels in the standard sentiment type, such as anger, sadness, etc., which are specifically set by those skilled in the art based on the historical public opinion assessment process, and will not be elaborated here.

[0048] The target topology map module is used to obtain a dynamic hierarchical target distribution map based on the topic key information space and the basic information space of the collected users, combined with the corresponding sentiment type, negative sentiment intensity, target severity score change rate and topic attention change rate at each level, through map layer algorithm combined with graph neural network and HSV color space.

[0049] The strategy module is used to generate target resolution strategies based on the dynamic hierarchical target distribution map, a pre-set deep target strategy library, and text generation algorithms. Based on the target resolution strategies, it monitors and analyzes in real time the target propagation rate, propagation range, target attention changes, and target severity scores of targets in the dynamic hierarchical target distribution map. Based on the analysis results, it dynamically adjusts the generated target resolution strategies until the target severity score of the target is lower than the pre-set target severity score threshold.

[0050] It should be further explained that the deep target strategy library in this embodiment was constructed by those skilled in the art based on historical target processing strategies of different types, combined with knowledge graph algorithms and graph database algorithms, and will not be elaborated further here;

[0051] It should be further explained that one specific implementation of the state adjustment generation target solution strategy in this embodiment is as follows:

[0052] First, based on the dynamic hierarchical target distribution map, the strategy module first uses a text matching algorithm to compare the target features presented in the map with scene templates in the deep target strategy library to select the appropriate strategy framework; the target features include, but are not limited to, target propagation rate, range, attention level, and target severity score.

[0053] Second, using a pre-trained BERT model, combined with the specific content of the target, user profiles, and propagation trends, strategy details are filled in to generate an initial target resolution strategy. It should be further explained that during strategy implementation, target data is collected through a real-time monitoring module, propagation trends are predicted using time series analysis algorithms, and the effectiveness of strategy implementation is evaluated using comparative analysis. If the target severity score is found to be below the preset target severity score threshold, then based on a reinforcement learning algorithm, the strategy parameters are dynamically adjusted or the strategy template is re-matched, using the target severity score threshold as the target, iterating until the target is effectively controlled. The preferred time series analysis algorithm is ARIMA.

[0054] It should be further explained that the topic key information space in this embodiment includes the target topic type, target text keywords, relationship attributes, target initial timestamp, propagation time length, target propagation range distribution, and target propagation rate; the basic information space of the collected users includes the geographical location, age, gender, historical target types, frequency of attention, corresponding sentiment type, and historical negative sentiment intensity of each type of target.

[0055] Further, please refer to Figure 2 The classification and acquisition module in this embodiment includes a data acquisition unit, a target discrimination unit, and a target classification unit;

[0056] The data acquisition unit is used to collect information data from different platforms in real time, and to use the IKAnalyzer word segmenter to perform text segmentation, text tagging and stop word removal on the collected information data to obtain a preprocessed initial collection text set.

[0057] The target discrimination unit is used to judge targets based on the preprocessed initial collected text set, combined with a target discrimination model constructed by a pre-trained word vector model, random forest algorithm and preset dynamic discrimination threshold, and obtain the real-time target topic text space;

[0058] The preset dynamic discrimination threshold is constructed by combining historical target information with the rate of change of target severity score, the rate of change of topic attention, target propagation rate, and propagation range per unit time through an association algorithm.

[0059] It should be further explained that the process of obtaining the real-time target topic text space in this embodiment includes:

[0060] Based on the preprocessed initial text set, each word in the text is converted into a fixed-dimensional vector using a pre-trained word vector model, obtaining the vector representation of the text. Based on the obtained text vector representation, a random forest algorithm is used to train the text vectors, constructing multiple decision trees. The majority voting principle is used to determine whether a text is a target, yielding a preliminary target judgment result. Based on historical target information, by statistically analyzing indicators such as the rate of change in target severity score, the rate of change in topic attention, the target propagation rate, and the propagation range per unit time in historical data, an association algorithm is used to mine the correlation between these indicators and target occurrence. Based on the mined correlation, combined with preset confidence and support thresholds, dynamic discrimination thresholds corresponding to different indicator combinations are determined. Based on the preliminary target judgment result obtained from the random forest algorithm and the determined dynamic discrimination thresholds, the preliminary judgment result is adjusted. If the score of the preliminary judgment result is higher than the dynamic discrimination threshold, it is judged as a target; otherwise, it is judged as a non-target, ultimately obtaining the real-time target topic text space. The pre-trained word vector model is preferably Word2Vec or BERT; the association algorithm is preferably the Apriori algorithm.

[0061] The target classification unit is used to obtain the real-time target classification topic text space by combining the text vector matrix obtained by the pre-trained word vector model with the topic type library and the real-time target topic text space through the matching algorithm, and to label the type targets in real time. The topic type library is constructed by combining the historical three-level target text type text space obtained by the hierarchical analysis algorithm with the distributed database and the graph algorithm.

[0062] It should be further explained that the type target acquisition process in this embodiment includes:

[0063] Based on the real-time target topic text space, texts are encoded using a BERT pre-trained word vector model, converting each text into a 768-dimensional vector representation to obtain a text vector matrix. Based on three years of historical target text data, an analytic hierarchy process (AHP) algorithm is used to hierarchically classify text types according to a three-level structure (domain-subdomain-specific topic, e.g., "social welfare-education-campus safety"), constructing a historical three-level target text type text space. Based on this historical three-level target text type text space, a graph algorithm is used to store each level of text type and its relationships in the form of a directed acyclic graph. Combined with a distributed database (e.g., HBase), efficient data storage and retrieval are achieved, constructing a topic type library. Based on the obtained text vector matrix and the constructed topic type library, a cosine similarity algorithm is used to calculate the similarity between text vectors and representative vectors of each type in the topic type library. Texts with a similarity greater than 0.7 are labeled as their corresponding type. If multiple similarities exceed a threshold, the type with the highest similarity is selected for labeling. Based on the labeling results, all texts in the real-time target topic text space are labeled with their real-time types, ultimately obtaining a target classification text set with type labels.

[0064] The topic type library in this embodiment includes, but is not limited to, social and livelihood issues, covering fields such as education, healthcare, employment, and housing, such as "adjustments to school district housing policies" and "popularization of new medical technologies"; entertainment and culture, including sub-fields such as film, music, games, and literature, such as "discussions sparked by the release of a popular movie" and "gameplay and reviews of a new game"; cutting-edge technology, involving fields such as artificial intelligence, blockchain, and new energy, such as "the application of artificial intelligence in medical image diagnosis" and "breakthroughs in the range technology of new energy vehicles"; and finance and economics, including stocks, funds, and macroeconomic policies, such as "the impact of central bank interest rate adjustments on the stock market" and "investment strategies and return analysis of a certain fund".

[0065] It should be further explained that the construction process of the topic type library in this embodiment includes:

[0066] Based on massive amounts of target text data from the past three years, this study employs a hierarchical analysis algorithm to meticulously classify text types according to a three-level structure of major category, intermediate category, and minor category, forming a historical three-level target text type text space, such as social welfare, education, and college admissions reform. Secondly, a graph algorithm is used to store the text types at each level and their relationships in the form of a directed acyclic graph, treating each type as a node and the parent-child relationships between types as directed edges. This is then combined with a distributed database, leveraging its high reliability, high performance, and scalability to store the graph structure data, achieving efficient data storage and retrieval, and ultimately constructing a topic type library.

[0067] Furthermore, the text extraction module in this embodiment includes a deep topic information extraction unit and a user information extraction unit;

[0068] The deep topic information extraction unit is used to extract target text keywords, relational attributes, propagation time length, corresponding target propagation range distribution, and target propagation rate obtained based on the target propagation range distribution and propagation time length in real time, according to the real-time target topic text space corresponding to each type of target and the first entity relation extraction model.

[0069] It should be further explained that one implementation method for information extraction by the first entity relation extraction model in this embodiment is as follows:

[0070] Based on the real-time target topic text space corresponding to each type of target, the text is encoded by pre-trained language models such as BERT, and each text is converted into a 768-dimensional vector representation containing contextual semantics to obtain the text semantic vector matrix.

[0071] Based on this matrix, a bidirectional long short-term memory network combined with a conditional random field algorithm is used to train a model for named entity recognition task. The model identifies entities in text and outputs a list of entities containing entity category and location information. The entities identified in the text include, but are not limited to, people, organizations, events, and locations.

[0072] Based on the identified entity list, a syntactic dependency parsing algorithm is used to analyze the dependency relationships between words in the sentence, extract the relationships between entities, and obtain a set of entity relationship pairs; the relationships between entities are exemplified as "company-product" and "policy-affected object";

[0073] Based on time-related keywords in the text and a time expression recognition algorithm, the start and end times of the target are extracted, and the propagation time length is calculated by the time difference; time-related keywords include, but are not limited to, "starts at", "continues", and "ends"; the preferred time expression recognition algorithm is the Dateparser library;

[0074] Based on geographic location keywords and place name recognition algorithms in the text, combined with map service APIs (such as Gaode Maps reverse geocoding), the keywords are converted into geographic coordinates, the coordinate distribution range is statistically analyzed, and the target propagation range distribution is obtained; the place name recognition algorithm is preferably GeoNames; geographic location keywords include, but are not limited to, city names and landmarks;

[0075] Based on the obtained target propagation range distribution and propagation time, the target propagation rate per unit time is calculated using the formula: Target propagation rate = Propagation range area / Propagation time.

[0076] Based on the extracted target text keywords, relational attributes, propagation time length, target propagation range distribution, and target propagation rate, structured deep topic information is integrated to complete the real-time extraction of key information for each type of target; the preferred syntactic dependency parsing algorithm is Stanford CoreNLP.

[0077] The user information extraction unit is used to extract the geographic location, age, gender, historical target types, frequency of attention, and comment text feature space of each user for each target type in real time based on the real-time target topic text space corresponding to each target type and the second entity relation extraction model.

[0078] It should be further explained that one implementation method for information extraction by the second entity relation extraction model in this embodiment is as follows:

[0079] Based on the real-time target topic text space corresponding to each type of target, the text is encoded using a Flair pre-trained model, converting each token in the text into a vector containing rich semantic features to obtain a token-level semantic vector sequence. Based on this vector sequence, a BiLSTM-CRF-based named entity recognition algorithm is trained on a pre-labeled user attribute dataset (containing entities related to geographic location, age, and gender) to identify entities such as the user's geographic location (e.g., city, region), age description (e.g., "25 years old", "youth"), and gender identifier (e.g., "he", "woman") in the text, and outputs a list of user attribute entities containing entity type and location.

[0080] Based on the user attribute entity list, a rule matching algorithm is used. For geographical location entities, regular expressions are used to match place name keywords (such as province, city, and county names). For age entities, numerical extraction and age group mapping rules (such as 18-25 years old as "youth") are used for standardization. For gender entities, a preset gender vocabulary (such as "male", "female", "Mr", "Ms") is used for confirmation to obtain standardized basic user attribute information.

[0081] Based on historical target data, the Word2Vec algorithm is used to train a corpus of users’ historical behavior to generate word vector representations of users’ historical interest types. The cosine similarity algorithm is used to calculate the similarity between the current target text vector and the historical type target vector. Historical type targets with similarity greater than a preset similarity threshold are taken as users’ historical interest types, and the corresponding interest frequency is counted.

[0082] Based on each user's comment text for each type of target, the TF-IDF algorithm is used to extract keywords from the comment text, and the TextRank algorithm is used to calculate keyword weights to construct a comment text feature vector containing high-frequency keywords and their weights. At the same time, the Doc2Vec algorithm is used to convert the comment text into a fixed-dimensional vector representation to obtain the comment text feature space. Based on the extracted user's geographical location, age, gender, historical target types, frequency of attention, and comment text feature space, a structured user information dataset is formed to complete the real-time extraction of user information.

[0083] Furthermore, the proactive evaluation module in this embodiment includes an emotion evaluation unit, an attention evaluation unit, a comprehensive evaluation unit, and a hierarchical correlation analysis unit;

[0084] The sentiment assessment unit is used to obtain the sentiment type of each user for each type of target based on the target text keywords, relational attributes, historical attention type targets, and the feature space of each user's comment text for each type of target, combined with the sentiment classification sub-model and sentiment type library in the pre-trained sentiment assessment model. Simultaneously, based on the assessment sub-model in the pre-trained sentiment assessment model, it obtains the negative sentiment intensity of each user for each type of target based on the attention frequency, propagation time length, corresponding target propagation range distribution, target propagation rate, and the keyword information and frequency of occurrence of sentiment-expressing information in each user's comment text feature space for each type of target. Finally, based on the attention sub-layer in the pre-trained sentiment assessment model and the negative sentiment intensity and type of each user for each type of target within a preset continuous time period, it obtains the negative sentiment intensity change sequence of each user for each type of target.

[0085] It should be further explained that the construction process of the sentiment classification sub-model in this embodiment is as follows:

[0086] Based on the target text keywords, relational attributes, historical attention types, and comment text feature space corresponding to each type of target, the text is encoded using a RoBERTa pre-trained language model, converting it into a 768-dimensional vector representation containing contextual semantics, thus obtaining a text semantic vector. Based on this vector, a classifier is constructed using a multi-layer fully connected neural network, fine-tuned on a large-scale labeled target sentiment dataset, and the model parameters are optimized using a cross-entropy loss function. The output is the probability distribution of each text belonging to different sentiment types, obtaining the sentiment type prediction result for each user for each type of target. The prediction result is matched with the standard types in the sentiment type library. If the maximum probability exceeds a threshold, it is determined to be the corresponding sentiment type; otherwise, it is marked as "uncertain," ultimately obtaining an accurate sentiment type determination. In this embodiment, the labeling includes four categories: positive, negative, neutral, and mixed.

[0087] It should be further explained that the construction process of the emotion type library in this embodiment is as follows:

[0088] Based on historical target text and manually annotated sentiment tags, a clustering algorithm is used to aggregate similar sentiment expressions to form sentiment category clusters. For each cluster, core keywords and semantic features are manually extracted to construct structured data containing sentiment type names, keyword sets, semantic descriptions, and example texts. This structured data is stored in a database, and a hierarchical relationship between types is established. For example, "negative" includes "anger" and "disappointment," forming a hierarchical sentiment type library to provide a standardized reference for sentiment classification. The preferred clustering algorithm is the K-Means algorithm; the keyword set includes, but is not limited to, "anger" corresponding to "rebuke" and "dissatisfaction"; the sentiment type names include, but are not limited to, "anger" and "support."

[0089] It should be further explained that the process of constructing the evaluation sub-model in this embodiment is as follows:

[0090] Based on the frequency of attention, duration of dissemination, distribution of target dissemination range, target dissemination rate, and sentiment index keywords in the comment text for each type of target, a normalization algorithm is used to scale numerical features (such as frequency of attention and target dissemination rate) to the [0,1] interval. Based on the normalized features, a regression model is constructed using the XGBoost gradient boosting tree algorithm. The model is trained on historical data labeled with negative sentiment intensity (1-5 points). The model is optimized using mean squared error as the loss function. The current target features are input, and the predicted value of negative sentiment intensity for each user for each type of target is output to obtain the final negative sentiment intensity score.

[0091] It should be further explained that the process of constructing the attention sublayer in this embodiment is as follows:

[0092] Based on the negative sentiment intensity and type sequence of each user for each type of target within a preset continuous time period, a two-dimensional feature matrix is ​​constructed by concatenating the intensity values ​​and type encoding (such as one-hot encoding). Based on this matrix, a multi-head self-attention mechanism is used to calculate the attention weights between features at different time steps. By weighted summation of the feature matrix, time steps that significantly influence sentiment intensity changes are highlighted, resulting in a weighted feature representation. Based on the weighted features, a fully connected layer is used for dimensionality reduction, outputting a one-dimensional trend value of negative sentiment intensity changes. The trend values ​​from multiple time steps are combined into a sequence to obtain the negative sentiment intensity change sequence of each user for each type of target within the preset time period. One-hot encoding is preferred for type encoding.

[0093] The attention evaluation unit is used to obtain the attention level of each user at each geographical location to each type of target at each time by combining the frequency of attention, the length of the propagation time, the distribution of the corresponding target propagation range and the target propagation rate of each user for each type of target with a fuzzy algorithm. It also obtains the continuous change rate sequence of the attention level of each user at each geographical location to each type of target within a preset continuous time length by the change rate of attention frequency, the distribution of the corresponding target propagation range and the target propagation rate.

[0094] It should be further explained that the specific process of implementing the fuzzy algorithm in this embodiment includes:

[0095] Based on each user's attention frequency, propagation time, corresponding target propagation range distribution, and target propagation rate for each type of target, a normalization algorithm is used to uniformly scale the numerical features such as attention frequency, propagation time, propagation range area, and target propagation rate to the [0,1] interval to obtain a standardized feature vector.

[0096] Based on standardized feature vectors, a fuzzy model is constructed by defining fuzzy sets and membership functions. The frequency of attention is divided into three fuzzy sets: low, medium, and high. A Gaussian function is used to describe the degree of membership. Similarly, the propagation time length, target propagation range distribution, and target propagation rate are defined with corresponding fuzzy sets and membership functions. The membership degree of each feature value to different fuzzy sets is calculated. The specific numerical division of the frequency of attention into three fuzzy sets can be determined by those skilled in the art according to their needs, and will not be elaborated here.

[0097] Based on the membership degree of each feature, fuzzy reasoning is performed by formulating a fuzzy rule base. For example, rules such as "if the frequency of attention is high and the target propagation rate is fast, then the degree of attention is high" are set. The Mamdani reasoning method is used, and the minimum value of the membership degree of the rule premise is taken as the membership degree of the conclusion.

[0098] Based on the fuzzy inference results, the centroid method is used to perform defuzzification calculations, converting fuzzy values ​​into precise values, and obtaining the degree of attention that users in each geographical location pay to each type of target at each time.

[0099] Based on attention data within a preset continuous time period, the rate of change of attention is obtained by calculating the ratio of the difference in attention at adjacent time points to the time interval. The rate of change at each time point is then arranged sequentially to obtain a continuous rate of change sequence of attention of users at each geographic location to each type of target within the preset continuous time period.

[0100] The comprehensive evaluation unit is used to obtain the target severity score of each type of target at each location at each time point based on the negative sentiment intensity of all users at each location at each time point and the attention level of all users at each geographic location at each time point for each type of target. It also obtains the target severity score change rate of each location for each type of target at a preset continuous time length based on the negative sentiment intensity change sequence and the continuous change rate sequence of attention level of all users at each geographic location for each type of target at a preset continuous time length, as well as the target severity score change rate of the same type of target at different locations at the same time point.

[0101] It should be further explained that the specific process of implementing the comprehensive evaluation unit in this embodiment includes:

[0102] Based on the data of negative sentiment intensity and attention level of all users at each location at each time point to each type of target, the principal component analysis algorithm is used to reduce the dimensionality of these two types of data, extract the principal components that best represent the data features, and obtain the dimensionality-reduced feature vector.

[0103] Based on the dimensionality-reduced eigenvectors, a judgment matrix is ​​constructed using the analytic hierarchy process (AHP). Domain experts are invited to score the relative importance of negative emotion intensity and attention level. By calculating the eigenvectors and the largest eigenvalue of the judgment matrix, the weight coefficients of the two are determined.

[0104] Based on the weighting coefficients, the intensity of negative emotions and the degree of attention are integrated and calculated using a weighted summation algorithm to obtain the target severity score for each type of target at each time point.

[0105] Based on the negative sentiment intensity change sequence and attention rate change sequence of all users for each type of target in each geographical location within a preset continuous time length, the time dimension of the two sequences is aligned by a dynamic time warping algorithm to make them comparable on a time scale.

[0106] Based on the aligned sequence, a mathematical model is established using a multiple linear regression algorithm to connect the negative emotion intensity change sequence, the continuous change rate of attention sequence, and the change of target severity score. The change rate of target severity score for each type of target at each position is calculated under a preset continuous time length.

[0107] Based on the target severity score data at different locations at the same time point, the rate of change of the target severity score of the same type of target at different locations at the same time point is obtained by calculating the difference and average of the target severity scores at each location.

[0108] The hierarchical association analysis unit is used to perform hierarchical association analysis based on the target severity score of each type of target at each location at each time point and the rate of change of the target severity score of each type of target at each location over a preset continuous time period, combined with hierarchical rules constructed according to the user's geographic location, age, gender, and target type of interest, to obtain the hierarchical target status association space; the hierarchical relationship corresponding to the hierarchical rules is: user's geographic location level > gender level > age segment level > target type of interest level.

[0109] As should be further explained, the specific process of implementing hierarchical association analysis in this embodiment includes:

[0110] Based on the target severity score of each target at each location and type at each time point and the rate of change of target severity score over a preset continuous time length, outliers are removed and missing values ​​are filled by data cleaning algorithms to obtain a standardized target severity score dataset.

[0111] Based on the user's geographic location, age, gender, and interest type, hierarchical rules are constructed. Following the priority order of "geographic location level > gender level > age segment level > interest type target level", the hierarchical weight matrix is ​​constructed using the analytic hierarchy process (AHP). The weight coefficients of each level are determined by calculating the eigenvectors of the matrix, thus obtaining the hierarchical rule weight vector.

[0112] Based on a standardized target severity score dataset, the data is grouped and aggregated according to geographic location level. The mean and rate of change of target severity scores for all users under each geographic location are calculated to obtain target status data classified by geographic location.

[0113] Based on the geographic location-based hierarchical data, the data is further grouped by gender level. Under each geographic location, the mean and rate of change of the target severity score for different genders are calculated to obtain target status data with superimposed gender levels.

[0114] Based on the above superimposed data, the target severity scores and change rates of users in different age groups are calculated under each geographical location and gender combination to obtain the target status data of superimposed age levels.

[0115] Based on the data of superimposed age levels, the data is finally grouped by the target level of the concern type. Under each combination of geographic location, gender and age segment, the mean and rate of change of the target severity score of users with different concern types are calculated to obtain complete hierarchical target status data.

[0116] Based on complete hierarchical target state data, an association rule mining algorithm (such as the Apriori algorithm) is used, with minimum support and confidence thresholds set, to mine the association relationships between target states at each level, such as "the association pattern between the types of targets that young women pay attention to and the changes in target severity scores under a certain geographical location"; the preferred association rule mining algorithm is the Apriori algorithm.

[0117] Based on the discovered relationships, graph database construction technology is used to store nodes and relationships at each level as a directed graph structure. Nodes include attributes such as geographic location, gender, age segment, interest type, and corresponding target severity score, while edges include the support and confidence attributes of association rules, thus obtaining a hierarchical target state association space.

[0118] It should be further noted that the target topology graph module in this embodiment includes a target graph construction unit, a target graph mapping unit, and a dynamic adjustment unit;

[0119] The target graph construction unit is used to define the initial position point of each type of target as the starting node, and each position point in the target propagation range distribution as the propagation branch node. At the same time point, the propagation connection relationship is constructed by multiplying the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point with the rate of change of the target severity score of the same type of target at different positions at the same time point. Based on the starting node, propagation branch node and propagation connection relationship, combined with the graph algorithm, the initial target propagation node graph is constructed.

[0120] It needs further explanation that the specific process of implementing the target graph construction unit in this embodiment includes:

[0121] Based on the initial location point of each type of target and the location points in the target propagation range distribution, each location point is converted into a node object containing latitude and longitude information through a unique geographic coordinate encoding algorithm. The node ID adopts the format of "location type_time stamp_serial number" to obtain the set of starting node and propagation branch nodes.

[0122] Based on the target propagation rate from the starting node to the propagation branch node and between the branch nodes at the same time point, and the rate of change of the target severity score of the same type of target at different locations, the product of the two is calculated by matrix multiplication, and then the Min-Max normalization algorithm is used to scale the product value to the [0,1] interval to obtain the edge weight matrix.

[0123] Based on the node set and edge weight matrix, an algorithm for constructing a directed graph adjacency list is used to store the neighboring nodes that each node points to and the corresponding edge weights, while also recording the geographical location attributes of the nodes, thus constructing an initial target propagation directed graph.

[0124] Based on the spatial distribution characteristics of nodes in the graph structure, the spatial connection relationship between nodes is generated by the Delaunay triangulation algorithm as a supplementary edge set. The edge weight matrix is ​​then used to filter edges with weights higher than the threshold of 0.3 to optimize the connectivity of the graph structure.

[0125] Based on the optimized graph structure, a topological sorting algorithm is used to sort the nodes according to the priority of the propagation path, generating a directed acyclic graph containing the hierarchical relationship of the nodes, and finally obtaining a structured initial target propagation node graph.

[0126] The target graph mapping unit is used to construct a hierarchical target severity score color map containing a color mapping function based on the target severity score at each level in the hierarchical target state association space and the HSV space using a support vector machine. Simultaneously, it constructs a propagation color map using the product of the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point, and the rate of change of the target severity score corresponding to the same type of target at different locations at the same time point, and the HSV space. This hierarchical target severity score color map is then mapped to each node in the initial target propagation node graph, and simultaneously mapped to the corresponding propagation connection relationship, thus constructing a hierarchical target distribution map. The hierarchical target distribution map includes the target severity scores corresponding to the user's geographic location level, gender level, age segment level, and interest type target level, and is distributed from top to bottom according to the level size.

[0127] It needs further explanation that the specific process of the target graph mapping unit implemented in this embodiment includes:

[0128] Based on the target severity score data at each level in the hierarchical target state association space, the target severity score values ​​are mapped to the [0,1] interval through a data normalization algorithm to obtain a standardized target severity score sequence.

[0129] Based on the standardized target severity score sequence and HSV color space, a support vector machine algorithm is used to train a color mapping function with the target severity score as the feature vector and the preset HSV color parameters as the labels. The HSV color parameters are: hue H∈[0,360], saturation S∈[0,1], and brightness V∈[0,1].

[0130] Based on the target propagation rate from the starting node to the propagation branch node and between the branch nodes at the same time point, and the product data of the target severity score change rate of the same type of target at different locations, the product value is divided into 5 intervals (low, lower, medium, higher, and high) through a data discretization algorithm to obtain the discretized propagation intensity level; furthermore, the details of dividing the product value into 5 intervals in this embodiment are specifically set by those skilled in the art based on the specific historical division process, and will not be described in detail here;

[0131] Based on discretized propagation intensity levels and the HSV color space, a propagation color mapping table is constructed by assigning different HSV color parameters to each level using a linear interpolation algorithm; for example, low levels correspond to low saturation and high brightness, and high levels correspond to high saturation and low brightness, etc.

[0132] Based on the nodes in the initial target propagation node graph, the target severity score corresponding to each node is input into the color mapping function through a traversal algorithm to obtain HSV color parameters and convert them into RGB color values ​​to complete the node color mapping;

[0133] Based on the propagation connection relationship in the initial target propagation node graph, the propagation color mapping table is matched by the edge weight (the product of the target propagation rate and the rate of change of the target severity score) through the search algorithm, the corresponding HSV color parameters are obtained and converted into RGB color values, and the edge color mapping is completed.

[0134] Based on the color mapping results of nodes and edges, the node groups of different levels are arranged from top to bottom in the order of "geographical location level > gender level > age segmentation level > target level of interest type" using a hierarchical layout algorithm. The node spacing and edge display hierarchy are adjusted to finally construct a hierarchical target distribution map.

[0135] The dynamic adjustment unit is used to collect in real time the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point, the product of the rate of change of the target severity score of the same type of target at different locations at the same time point, and the location distribution of the real-time propagation range. Combining the propagation path algorithm and the spatial rectangle expansion algorithm, it adjusts the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationship between different nodes in real time laterally. At the same time, it uses the rate of change of the target severity score of each type of target at each location under a preset continuous time length to adjust the target severity score and the corresponding color mapping depth of each node corresponding to each level in the hierarchical target distribution map vertically, so as to obtain a dynamic hierarchical target distribution map.

[0136] It should be further explained that the process of real-time horizontal adjustment of the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationships between different nodes in this embodiment includes:

[0137] Based on the real-time collected target propagation rate and target severity score change rate, the data is normalized to the [0,1] interval using a normalization algorithm, and then the product of the two is calculated to obtain the weighted propagation connectivity relationship.

[0138] Based on the initial propagation node graph and historical data, the GraphSAGE graph neural network combined with the PageRank algorithm is used to predict potential propagation nodes and generate Top-K propagation paths;

[0139] Based on the real-time propagation range coordinates and potential propagation nodes, the range is initialized using the minimum enclosing rectangle algorithm. Then, the spatial decay weight is calculated by combining the Gaussian kernel function with the rate of change of target severity scores between different nodes. Combined with the expansion coefficient constructed by the target propagation rate per unit time, the dynamically expanded spatial rectangle range is obtained.

[0140] Based on the edge weights corresponding to the propagation connection relationship and the HSV color space, the weights are converted into saturation and brightness through a linear mapping algorithm, and then the color depth of the propagation connection relationship is obtained through an HSV to RGB conversion algorithm.

[0141] Based on the adjusted rectangular range and color parameters, tiles are loaded via the map API and nodes and edges are rendered using WebGL hardware acceleration.

[0142] Based on the latest data and user interactions, the expansion coefficient and the edge weights corresponding to the propagation connection relationship are dynamically adjusted through reinforcement learning algorithms to obtain a horizontally updated hierarchical target distribution map.

[0143] It should be further explained that the specific process of vertically adjusting the hierarchical target distribution map in this embodiment includes:

[0144] Based on the rate of change of target severity score for each type of target at each location under a preset continuous time length, missing data points are filled by linear interpolation algorithm, and the Z-Score standardization algorithm is used to convert the sequence into a standard normal distribution with a mean of 0 and a standard deviation of 1, so as to obtain the preprocessed standardized rate of change sequence.

[0145] Based on this sequence, a multi-layer LSTM neural network combined with a seasonal decomposition algorithm is used to perform trend decomposition on the sequence, predict the rate of change at the next three time points, and obtain time series analysis results containing trend terms, periodic terms, and predicted values.

[0146] Based on the target severity score of the current node, the historical rate of change sequence is analyzed by an exponential decay weighted algorithm to calculate the historical impact value, and the current target severity score is updated by weighting the mean of the future predicted values ​​to obtain the adjusted target severity score.

[0147] Based on the adjusted target severity score, and referring to the historical target severity score quantiles, the levels are divided into 5 levels. Using the CIE Lab color space conversion algorithm, the levels are mapped to a color parameter coordinate system with brightness from 80 to 30, red-green axis from -20 to 60, and yellow-blue axis from 20 to 0. Then, it is converted to the HSV color space, so that the saturation changes with the level to obtain the corresponding HSV color value.

[0148] Based on HSV color values, the color values ​​are applied to nodes of the corresponding level through the node style interface of the map rendering engine. A linear interpolation algorithm is used to achieve a smooth 500ms transition animation of color changes. At the same time, the node radius is adjusted according to the target severity score. For example, the base radius is 5 pixels, and 1 pixel is added for every 10% of the target severity score. Finally, the target severity score and color mapping depth are updated in the vertical dimension.

[0149] This embodiment integrates multi-module collaboration and deep learning technology to construct a full-process intelligent system covering target acquisition, analysis, visualization, and strategy generation. It realizes a technological upgrade of target management from passive response to proactive intervention, significantly improves the accuracy and timeliness of target processing, and provides systematic technical support for target prevention and control in multiple scenarios.

[0150] At the data acquisition and preprocessing level, the classification acquisition module achieves efficient information filtering and standardization through real-time multi-platform capture and IKAnalyzer word segmentation technology. The data acquisition unit performs word segmentation, tagging, and stop word removal on the original text, effectively reducing data noise and providing a high-quality text foundation for subsequent analysis. The target discrimination unit combines the Word2Vec / BERT pre-trained word vector model and the random forest algorithm to convert the text into semantic vectors and then performs preliminary discrimination through a multi-decision tree voting mechanism. At the same time, it mines the correlation between historical target indicators based on the Apriori association algorithm and constructs a dynamic discrimination threshold. Through the dual mechanism of model prediction and threshold adjustment, the accuracy of target recognition is greatly improved, avoiding the missed and false judgment problems of traditional fixed threshold methods. The target classification unit uses BERT encoding to generate a high-dimensional text vector matrix. Combined with the three-level structure of the hierarchical topic type library and the cosine similarity matching algorithm, it achieves real-time classification and labeling of target text, providing a structured data foundation for subsequent topic tracking and user profile construction, enabling targets in different fields (including but not limited to social welfare, cutting-edge technology, etc.) to be quickly located and classified.

[0151] The text extraction module achieves in-depth mining of target information and user profiles through multi-model fusion; the deep topic information extraction unit, based on BERT semantic encoding and BiLSTM-CRF named entity recognition algorithm, combined with Stanford CoreNLP syntactic dependency analysis, not only achieves accurate identification of entities such as people, organizations, and locations, but also extracts complex relationship attributes such as policy-impacted objects; it uses the Dateparser library and map service API to extract the temporal and spatial features of target propagation, and calculates the target propagation rate through formulas, thus fully outlining the spatiotemporal context of target propagation, making the target diffusion process quantifiable and traceable; the user information extraction unit uses the Flair pre-trained model and BiLSTM-CRF algorithm to identify basic user attributes, trains a historical behavior corpus through Word2Vec and combines cosine similarity to analyze user attention preferences, and uses TF-IDF and TextRank algorithms to construct the feature space of comment text, realizing the construction of multi-dimensional user profiles from basic attributes to sentiment tendencies, providing accurate user dimension data support for subsequent sentiment analysis and strategy generation, enabling the system to conduct differentiated analysis of target preferences for user groups with different characteristics.

[0152] The proactive assessment module achieves deep insight into the target situation through multi-dimensional quantitative analysis. The sentiment assessment unit constructs a sentiment classification sub-model based on a RoBERTa pre-trained model and a multi-layer fully connected neural network, combining it with a hierarchical sentiment type library generated by K-Means clustering to achieve accurate sentiment type determination. It utilizes an XGBoost regression model to fuse multi-dimensional features with sentiment keywords from comment text to quantify the intensity of negative sentiment. Through a multi-head self-attention mechanism, it captures the dynamic changes in sentiment intensity over time, constructing a negative sentiment intensity change sequence, achieving spatiotemporal dual-axis analysis of the sentiment dimension, enabling precise tracking of the target sentiment evolution process. The attention assessment unit uses a fuzzy algorithm to standardize multi-source features, then defuzzifies them using Gaussian membership functions, Mamdani inference, and the centroid method to quantify different time-series features. The empty user attention level and change rate solve the nonlinear mapping problem of multi-indicator fusion; the comprehensive evaluation unit determines the weights through principal component analysis dimensionality reduction and hierarchical analysis, calculates the target severity score by weighted fusion of emotional intensity and attention level, and generates the target severity score change rate by combining dynamic time warping and multiple linear regression, realizing a quantitative assessment of the target status; the hierarchical association analysis unit constructs a graph structure containing nodes and relationships at each level according to the hierarchical rule of "geographic location > gender > age > attention type", and reveals the inherent association pattern between different user groups and target status through data cleaning, grouping aggregation and Apriori association mining, providing a multi-perspective analysis dimension for target tracing and precise intervention.

[0153] The target topology graph module uses graph algorithms and visualization techniques to achieve an intuitive presentation and dynamic update of the target situation. The target graph construction unit uses geographic coordinate encoding and a directed graph adjacency list to construct a propagation node graph. Combined with Delaunay triangulation to optimize graph connectivity, it generates a directed acyclic graph containing propagation levels through topological sorting, intuitively displaying the target's diffusion path from the starting node to branch nodes, thus visualizing the propagation network structure. The target graph mapping unit constructs an HSV color mapping function based on support vector machines, converting abstract target severity scores and propagation intensity into visually discernible color parameters. A hierarchical layout algorithm arranges node groups according to priority, achieving a visual mapping of multi-dimensional target information and enhancing the intuitive representation of the data. The dynamic adjustment unit predicts potential propagation nodes using GraphSAGE graph neural networks and the PageRank algorithm, constructs a spatial expansion model using Gaussian kernel functions, and updates the target severity score using LSTM and seasonal decomposition algorithms. This achieves horizontal expansion and vertical depth adjustment of the distribution map, ensuring that the visualization results reflect the target's evolution trend in real time, helping users quickly grasp both global and detailed changes in the target.

[0154] The strategy module constructs a closed-loop strategy system based on a dynamic hierarchical target distribution map, featuring intelligent generation, real-time evaluation, and adaptive optimization. By comparing text matching algorithms with scenario templates in a deep target strategy library, it quickly selects strategy frameworks that match target features (target propagation rate, target severity score, etc.), ensuring the initial targeting of the strategy. It uses a pre-trained BERT model combined with target content, user profiles, and propagation trends to fill in strategy details, and leverages natural language generation technology to enhance the readability and operability of the strategy text. During strategy implementation, the ARIMA time series algorithm predicts propagation trends, comparative analysis evaluates implementation effectiveness, and a feedback mechanism is built based on reinforcement learning algorithms to dynamically adjust strategy parameters or re-match strategy templates. This allows the strategy to continuously iterate as the target evolves until it falls below the risk threshold, significantly improving the intelligence level and actual effectiveness of target intervention and forming a complete technology chain from data-driven to decision optimization.

[0155] In summary, this embodiment achieves high efficiency, accuracy, and dynamism in target processing through the organic integration of deep learning algorithms, graph neural networks, and natural language processing technologies. It can not only quickly identify and classify target information and deeply mine target propagation characteristics and user profiles, but also intuitively present the target situation through visualization technology, ultimately generating adaptive and optimized solutions. This provides strong technical support for governments, enterprises, and other entities in scenarios such as public opinion guidance, risk prevention and control, and public event management, effectively improving the efficiency, scientific nature, and foresight of target management.

[0156] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

[0157] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A text processing system based on topological analysis, characterized in that, include: The module includes a classification and data collection module, a text extraction module, an active evaluation module, and a target topology map module. The classification and acquisition module is used to collect online text information in real time and combine it with a preset target discrimination model to obtain a real-time target topic text space, and to obtain a real-time target classified topic text space based on the real-time target topic text space and a topic type library; the target represents public opinion information. The text extraction module extracts key information space of the topic based on the real-time target classification topic text space, and simultaneously extracts basic information space of the collected users based on the real-time target classification topic text space. The key information space of the topic includes the target topic type, target text keywords, relationship attributes, target initial timestamp, propagation time length, target propagation range distribution, and target propagation rate. The basic information space of the collected users includes the geographical location, age, gender, historical target types, and frequency of interest. The active evaluation module is used to obtain the target severity score, target severity score change rate, and topic attention change rate for each level of the target based on the topic key information space and the basic information space of the collected users. The target severity score is obtained by combining the negative emotional intensity of all users at each location at each time point for each type of target with the attention level of all users at each geographical location for each type of target at each time point. The target topology map module is used to: define the initial position point of each type of target as the starting node, each position point in the target propagation range distribution as the propagation branch node, and construct the initial target propagation node map by using the product of the target propagation rate from the starting node to each propagation branch node and between propagation branch nodes at the same time point and the rate of change of the target severity score of the same type of target at different positions as the propagation connection relationship; construct a hierarchical target severity score color map containing a color mapping function based on the target severity score at each level in the hierarchical target state association space and the HSV space, and construct a propagation color map using the size of the product and the HSV space, and map the hierarchical target severity score color map to each node in the initial target propagation node map, and map the propagation color map to the corresponding propagation connection relationship to construct a hierarchical target distribution map; real-time acquisition The distribution of the product and propagation range is used to horizontally adjust the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationships between different nodes. Simultaneously, using the rate of change of target severity scores for each type of target at each location over a preset continuous time period, the target severity scores and corresponding color depths at each level under each node in the hierarchical target distribution map are vertically adjusted to obtain a dynamic hierarchical target distribution map. Each level is divided according to the priority order of the user's geographical location, gender, age segment, and target type of interest. The hierarchical target state association space is specifically defined as follows: the target severity scores and the rate of change of scores are standardized; according to the priority order, the weight of each level is determined using the analytic hierarchy process (AHP); the mean score and rate of change are aggregated level by level; the Apriori algorithm is used to mine the association relationships between each level; and the data is stored in a directed graph structure in a graph database.

2. The text processing system based on topology analysis as described in claim 1, characterized in that, The basic information space of the collected users also includes the historical evaluation of the negative emotional intensity of each type of target by the collected users; the classification collection module includes a data collection unit and a target discrimination unit; The data acquisition unit is used to collect information data from different platforms in real time, and to use the IKAnalyzer word segmenter to perform text segmentation, text tagging and stop word removal on the collected information data to obtain a preprocessed initial collection text set. The target discrimination unit is used to judge targets based on the pre-processed initial collected text set, combined with a target discrimination model constructed by a pre-trained word vector model, a random forest algorithm, and a preset dynamic discrimination threshold, to obtain a real-time target topic text space; the preset dynamic discrimination threshold is constructed by combining historical target information with the target severity score change rate, topic attention change rate, target propagation rate, and propagation range per unit time through an association algorithm.

3. The text processing system based on topology analysis as described in claim 2, characterized in that, The text extraction module includes a deep topic information extraction unit; the deep topic information extraction unit is used to extract the target text keywords, relational attributes, propagation time length, corresponding target propagation range distribution, and target propagation rate obtained based on the target propagation range distribution and propagation time length in real time according to the real-time target topic text space corresponding to each type of target and the first entity relation extraction model.

4. The text processing system based on topology analysis as described in claim 3, characterized in that, The text extraction module also includes a user information extraction unit; the active evaluation module includes a sentiment evaluation unit. The user information extraction unit is used to extract the geographical location, age, gender, historical target types, frequency of attention, and comment text feature space of each user for each target type in real time based on the real-time target topic text space corresponding to each target type and the second entity relation extraction model. The sentiment assessment unit is used to obtain the sentiment type of each user for each type of target based on the target text keywords, relational attributes, historical attention type targets, and the feature space of each user's comment text for each type of target, combined with the sentiment classification sub-model and sentiment type library in the pre-trained sentiment assessment model. At the same time, based on the assessment sub-model in the pre-trained sentiment assessment model, it obtains the negative sentiment intensity of each user for each type of target based on the attention frequency, propagation time length, corresponding target propagation range distribution, target propagation rate, and the keyword information and frequency of occurrence of the sentiment degree expressed in the feature space of each user's comment text for each type of target. Based on the attention sublayer in the pre-trained emotion assessment model, combined with the negative emotion intensity and type of each user towards each type of target within a preset continuous time period, the sequence of changes in negative emotion intensity for each user towards each type of target is obtained.

5. The text processing system based on topology analysis as described in claim 4, characterized in that, The proactive evaluation module further includes an attention evaluation unit and a comprehensive evaluation unit. The attention evaluation unit is used to obtain the degree of attention of users at each geographical location to each type of target at each time point based on the frequency of attention, duration of propagation, distribution of the corresponding target propagation range, and target propagation rate of each user for each type of target, combined with a fuzzy algorithm. It also obtains a continuous change rate sequence of the degree of attention of users at each geographical location to each type of target within a preset continuous time period by using the change rate of attention frequency, distribution of the corresponding target propagation range, and target propagation rate. The comprehensive evaluation unit is used to obtain the target severity score of each type of target at each location at each time point. Based on the change sequence of negative sentiment intensity and the continuous change rate sequence of attention of all users at each geographical location to each type of target within a preset continuous time period, it obtains the change rate of the target severity score of each type of target at each location within a preset continuous time period and the change rate of the target severity score of the same type of target at different locations at the same time point.

6. The text processing system based on topology analysis as described in claim 5, characterized in that, The proactive assessment module also includes a hierarchical correlation analysis unit; the hierarchical correlation analysis unit is used to obtain a hierarchical target status correlation space based on the target severity score of each type of target at each location at each time point and the rate of change of the target severity score of each type of target at each location over a preset continuous time period, combined with the hierarchical rules constructed based on the user's geographical location, age, gender, and target type of interest; the hierarchical relationship corresponding to the hierarchical rules is: user's geographical location level > gender level > age segment level > target type of interest level.

7. The text processing system based on topology analysis as described in claim 6, characterized in that, The hierarchical target distribution map includes target severity scores corresponding to the user's geographical location level, gender level, age segment level, and target level of interest type, and is distributed from top to bottom according to the level.

8. The text processing system based on topology analysis as described in claim 7, characterized in that, The process of horizontally adjusting the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationships between different nodes includes: Based on the real-time collected target propagation rate and target severity score change rate, the data is normalized to the [0,1] interval using a normalization algorithm, and then the product of the two is calculated to obtain the weighted propagation connectivity relationship. Based on the initial propagation node graph and historical data, the GraphSAGE graph neural network combined with the PageRank algorithm is used to predict potential propagation nodes and generate Top-K propagation paths; Based on the real-time propagation range coordinates and potential propagation nodes, the range is initialized using the minimum enclosing rectangle algorithm. Then, the spatial decay weight is calculated by combining the Gaussian kernel function with the rate of change of target severity scores between different nodes. Combined with the expansion coefficient constructed by the target propagation rate per unit time, the dynamically expanded spatial rectangle range is obtained.

9. The text processing system based on topology analysis as described in claim 8, characterized in that, The process of horizontally adjusting the distribution range of the hierarchical target distribution map and the color depth of the corresponding connection relationships between different nodes also includes: Based on the edge weights corresponding to the propagation connection relationship and the HSV color space, the weights are converted into saturation and brightness through a linear mapping algorithm, and then the color depth of the propagation connection relationship is obtained through an HSV to RGB conversion algorithm. Based on the adjusted rectangular range and color parameters, tiles are loaded via the map API and nodes and edges are rendered using WebGL hardware acceleration. Based on the latest data and user interactions, the expansion coefficient and the edge weights corresponding to the propagation connection relationship are dynamically adjusted through reinforcement learning algorithms to obtain a horizontally updated hierarchical target distribution map.

Images