A personalized topic text public opinion material generation method and system

Abstract

The application belongs to the technical field of network space cognition domain, and discloses a personalized topic text public opinion material generation method and system. The method introduces influential user discovery under a topic into personalized text content generation, adopts a graph neural network method of graph link prediction and node importance to discover important users, automatically discovers high-value and important influential users from a sparse user social relationship network under a topic background, and takes the users as use objects of personalized text content generation. Meanwhile, in combination with a large-scale dialogue language model based on deep learning, the method fuses feature coding representation of social importance of the users and positive and negative sentiment of text, and automatically generates sentiment controllable personalized text based on social characteristics of the users. The application introduces influential user discovery under a certain topic and sentiment positive and negative personalized text generation, and solves the problem that review and dialogue short text generation cannot be accurately based on high-value social user attribute information under a context background.

Description

Technical Field

[0001] This invention belongs to the field of cyberspace cognitive domain technology, and in particular relates to a method and system for generating personalized topic text materials. Background Technology

[0002] Existing technologies include complex network analysis methods, text generation techniques, link prediction techniques, and node importance discovery techniques.

[0003] Complex network analysis methods, text generation techniques, link prediction techniques, and node importance discovery techniques all possess excellent algorithms in their respective fields, but there is a lack of application solutions that combine these techniques. In text generation applications, sometimes the generated text content does not take into account the user's social relationships and emotional state. The generated text may satisfy the context, but it fails to deeply reflect the user's stance and emotions, resulting in text that cannot intuitively and accurately express the user's viewpoint.

[0004] In summary, the problems and defects of the existing technology are as follows: (1) The text generation method using the content similarity matching technology of the text library and the general language model usually only considers the semantic logic level of the text context and lacks other knowledge feature information, resulting in inaccurate text generation.

[0005] (2) In the context of a specific social relationship, different users have different positions and emotions in their understanding of the same topic. Existing technological inventions do not take into account the attribute information and social characteristics of users in the social relationship network when creating text content, and cannot generate texts that are personalized and accurately close to the user's context. Summary of the Invention

[0006] To overcome the problems existing in related technologies, the present invention discloses a method and system for generating personalized topic text materials.

[0007] The invention is as follows: A method for generating personalized topic-based text content. This method incorporates the discovery of influential users under a topic into the generation of personalized text content. It uses graph link prediction and node importance graph neural network methods to discover important users, automatically identifying high-value and influential users from the sparse social relationship network of users in the topic background, and using them as the target audience for personalized text content generation. At the same time, it combines a large-scale dialogue language model based on deep learning, integrates the feature encoding representation of the user's social importance and the positive or negative sentiment of the text, and automatically generates sentiment-controllable personalized text based on the user's social characteristics.

[0008] Specifically, the following steps are included:

[0009] S1, Building User Social Relationship Networks;

[0010] S2, predicting potential relationships in a user's social network based on graph links;

[0011] S3, Discovering High Social Value Targets;

[0012] S4 generates personalized text content by incorporating important user attribute information.

[0013] Furthermore, in step S1, the construction of the user's social relationship network includes:

[0014] Data related to current events and topics is obtained from social media platforms, including post and comment text data, user likes, comments, and reposts interaction data, and basic information about users' friend lists and following lists. This topic-related data is stored in a database. Then, using complex network analysis methods, a social relationship network with users as nodes is constructed using users' interaction data and friend and following lists.

[0015] Furthermore, by regularly updating the topic and event data obtained from social media platforms within the current time period, including topic text data and user interaction relationship data, the new data is simultaneously integrated into the user's social relationship network.

[0016] Furthermore, in step S2, predicting potential relationships in a user's social network based on graph links includes: using the PLNLP algorithm to predict links; based on the constructed user social network, using a negative sample sampling strategy to sample positive and negative samples, where positive samples are node pairs with existing connections in the network, and negative samples are node pairs without existing connections; and then encoding the neighborhood subgraphs of the node pairs using a graph neural network encoder to extract useful relationship information features from the network.

[0017] The node pair encoding represents the probability of two nodes being connected by an edge calculated using the link scoring function in the link predictor. Through the above calculation process, a batch of node pairs most likely to have edge relationships are predicted in the user social relationship network already constructed under the topic context.

[0018] Furthermore, the parameter learning of the entire link prediction model is achieved through the objective function of Pairwise Ranking. Specifically, this includes:

[0019] The link prediction model framework includes a neighborhood encoder, a link predictor, negative sampling, and a pairwise ranking objective function. The goal of link prediction is to rank positive samples (node ​​pairs with edges) higher than negative samples (node ​​pairs without edges). The pairwise ranking objective function uses the idea of ​​pairwise learning to optimize the parameters of the link prediction model, which is consistent with the overall goal of link prediction.

[0020] Furthermore, in step S3, the discovery of high social value targets includes: using a node importance recognition algorithm based on graph convolutional neural networks to discover key nodes in the network as users for personalized content generation; using finite-order neighborhood information of nodes in the network topology to learn low-dimensional vector representations of nodes to train the model; predicting the expanded user social relationship network based on links; using a random walk algorithm to calculate the neighborhood network of each node in the social relationship network; generating initial embedding vector representations of each node based on neighborhood network structure information; obtaining deep embedding representations of nodes through convolutional layers; and then using fully connected layers to predict the influence of each node in the user social relationship network.

[0021] Furthermore, the personalized text content generation in step S3, which incorporates important user attribute information, includes:

[0022] The dialogue content, relevant attribute information of the identified important users, and positive and negative sentiment information are used as input data. The input data is processed by a personalized text generation model to generate several emotionally controllable text contents that conform to the user's context. The autoencoder, neighborhood interaction encoder, and sentiment attention encoder in the personalized text generation model are used to encode the dialogue content, social relationship attributes of important users, and positive and negative sentiment information, respectively, to obtain the corresponding encoded feature vectors.

[0023] Based on the feature vector fusion module in the personalized text generation model, the encoded feature vectors output by each encoder are fused to obtain a fused feature vector; the decoder in the personalized text generation model is used to decode the fused feature vector; after discovering target users with high social value, positive and negative sentiment text content is generated based on the social relationship between the target user and their interactive friends and a large-scale dialogue language model.

[0024] Furthermore, among the generated text content with controllable emotions that conform to the user's context, the relevant attributes of important users have been identified, including the user's friend relationships and social relationships with users who have interacted with the user. The emotional positive and negative information includes two types of text emotional tendencies: positive and negative.

[0025] Furthermore, by combining the neighborhood interaction encoder that the model has already learned about users and their interaction relationships, during inference, only the input dialogue content and the positive and negative emotional information need to be encoded. The feature fusion module integrates the trained neighborhood interaction encoder and, after passing through the model decoder, generates several candidate text contents with controllable emotions that conform to the user's context.

[0026] Another object of the present invention is to provide a personalized topic text content generation system, the system comprising:

[0027] The user social relationship network construction module is used to obtain topic-related data from social media platforms within the current time period and store the topic-related data in the information database; then, through complex network analysis methods, it uses user interaction data and friend and following lists to construct a social relationship network with users as nodes.

[0028] The graph link prediction module for potential relationships in user social networks is used to predict the most likely pair of nodes to have edge relationships in the user social relationship network that has been constructed under the context of a topic.

[0029] The high social value target discovery module uses a node importance recognition algorithm based on graph convolutional neural networks to discover key nodes in the network as users for personalized content generation. It learns low-dimensional vector representations of nodes by utilizing finite-order neighborhood information in the network topology to train the model. Based on link prediction, it expands the user's social relationship network and uses a random walk algorithm to calculate the neighborhood network of each node in the user's social relationship network. It generates initial embedding vector representations of each node based on neighborhood network structure information. It obtains deep embedding representations of nodes through convolutional layers and then uses fully connected layers to predict the influence of each node in the user's social relationship network.

[0030] The personalized text content generation module, which incorporates important user attribute information, is used to generate sentiment-positive and negative text content based on the social relationships between the target user and their interacting friends, as well as a large-scale dialogue language model, after identifying target users with high social value.

[0031] Combining all the above-mentioned technical inventions, the advantages and positive effects of this invention are as follows: Based on the problems existing in the existing text content generation methods, this invention proposes a personalized topic text content generation method for important users on social networks. It introduces the discovery of influential users under a certain topic and the generation of personalized text with positive and negative emotions, solving the problem that it is impossible to accurately generate short texts of comments and conversations based on the attribute information of high-value social users in the context of the context.

[0032] This invention innovatively incorporates the discovery of influential users within a given topic into personalized text content generation. It employs graph neural network methods such as graph link prediction and node importance to discover important users, automatically identifying high-value and influential users from the sparse social relationship networks of users within the topic context, and using them as the target audience for personalized text content generation. Simultaneously, it combines a large-scale dialogue language model based on deep learning, considering the user's social importance and the positive or negative sentiment of the text, to automatically generate personalized text tailored to the user's social characteristics.

[0033] Compared with the prior art, the advantages of the present invention further include: in the field of topic guidance, the present invention can be used to automatically generate emotionally controllable text materials that conform to the user's context on social media platforms, providing reference materials for public opinion guidance and assisting in the effectiveness of accurate public opinion guidance; among the existing methods for generating related dialogue texts, there is no solution to introduce the user's social characteristics and positive or negative emotions into the generation of dialogue texts. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure;

[0035] Figure 1 This is a diagram of the personalized topic text content generation method and system provided in the embodiments of the present invention;

[0036] Figure 2 This is a flowchart of topic data preprocessing and user social relationship network construction provided in this embodiment of the invention;

[0037] Figure 3 This is an overall calculation flowchart for link prediction provided in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the node importance recognition framework based on graph convolutional neural networks provided in this embodiment of the invention.

[0039] Figure 5 This is a flowchart of a personalized text generation training process that incorporates important user attribute information, provided in an embodiment of the present invention.

[0040] Figure 6 This is a flowchart of personalized text generation reasoning that combines important user attribute information, provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0042] The innovation of the personalized topic text content generation method and system provided in this invention lies in the fact that by introducing factors such as user social attributes and positive or negative emotions, the generation of text content that is closer to the context is guided. It is necessary to first discover influential and important social users and integrate the user's social characteristics into the feature encoding representation to guide the generation of text.

[0043] Specifically, this invention creatively incorporates the discovery of influential users and their positive or negative sentiments under a certain topic into personalized text generation. It uses graph neural network methods such as link prediction and node importance discovery to discover high-value users, and encodes user social interaction characteristics and positive or negative sentiments through feature fusion. This solves the problem of not being able to accurately generate short texts for comments and conversations based on the attribute information of high-value social users in the context of context.

[0044] Example 1: The personalized topic text content generation method provided in this embodiment of the invention is based on target user discovery and large-scale dialogue language model text generation. It calculates and mines users with high influence and distinctive characteristics by using the social relationships of users in the context of the current topic, and then generates targeted high-quality comments or dialogue short text content through text generation methods.

[0045] Specifically, such as Figure 1 As shown, it includes the following steps:

[0046] S1, User Social Network Construction. Data related to current-time events is acquired from social media platforms, including topic text data such as posts and comments, user interaction data such as likes, comments, and shares, and basic information such as user friend lists and following lists. This topic-related data is stored in a database. Then, using complex network analysis methods, a social network with users as nodes is constructed based on user interaction data and friend / following lists. The entire process of acquiring topic event data and constructing user social networks is as follows: Figure 2 As shown. To ensure the comprehensiveness and timeliness of the tracking of topic events, it is necessary to update the topic event data regularly. The updated data includes topic text data and user interaction relationship data, and the new data is simultaneously updated and integrated into the user's social relationship network.

[0047] It is understandable that the purpose of complex network analysis methods is to abstract users as nodes, various relationships as edges, and construct social network structures.

[0048] The aforementioned periodic updating of topic event data refers to continuously adding newly generated topic data to the existing data set.

[0049] S2, Graph Link Prediction of Potential Relationships in User Social Networks. In the context of topic-related events, social networks built upon user interaction relationships are generally large in scale and sparsely connected, making it difficult to intuitively discover potential relationships or future interactions between users. Link prediction, based on the existing user social network structure, predicts the probability of a link forming between two unconnected nodes in the network, effectively supplementing user relationships hidden under topic interactions and assisting in identifying target users with specific characteristics within the topic's public opinion. This invention employs the PLNLP algorithm for link prediction. Based on the existing user social network, a negative sample sampling strategy is used to sample positive and negative samples. Positive samples are node pairs with existing edges in the network, while negative samples are node pairs without existing edges. A graph neural network encoder then encodes the neighborhood subgraphs of the node pairs to extract useful relational information features from the network. The node pair encoding represents the probability of two nodes being connected by a link scoring function in the link predictor; the parameter learning of the entire link prediction model is achieved through the objective function of PairwiseRanking. Through the above calculation process, a batch of node pairs most likely to have edge relationships in the existing user social relationship network under the topic context can be predicted (e.g., ...). Figure 3 (Middle dashed line), the computing architecture of this module is as follows: Figure 3 As shown.

[0050] It is understandable that the PLNLP algorithm consists of four components: neighborhood encoder, link predictor, negative sampling, and Pairwise Ranking objective function. This method is just a tool algorithm for discovering potential user relationships.

[0051] The PLNLP algorithm uses a link prediction model, which consists of four components: a neighborhood encoder, a link predictor, negative sampling, and a pairwise ranking objective function. The goal of link prediction is to ensure that the ranking of positive samples (node ​​pairs with edges) is higher than that of negative samples (node ​​pairs without edges). In order to maintain consistency with the overall goal of link prediction, the objective function adopts the idea of ​​pairwise learning ranking to optimize the parameters of the link prediction model.

[0052] A negative sample sampling strategy is used to sample positive and negative samples, and an adversarial learning approach is adopted for negative sample sampling.

[0053] Encoding the neighborhood subgraph of a node pair using a graph neural network encoder involves: obtaining the representation of each node in the neighborhood subgraph of the node pair through an attention mechanism using a graph attention neural network, and then obtaining the interaction feature representation of each pair of neighboring nodes through element-wise vector processing.

[0054] The node pair encoding representation is used to calculate the probability of two nodes being connected by a link scoring function in the link predictor. This includes: obtaining the feature representation of the node pair through a graph attention neural network, and calculating the probability of two nodes being connected by a link scoring function (such as a multilayer perceptron predictor or vector inner product calculation).

[0055] S3, High Social Value Target Discovery. Because topic and event data are updated periodically, the structure of the user social relationship network in link prediction changes. This invention employs a node importance recognition algorithm based on graph convolutional neural networks to discover key nodes in the network as users for personalized content generation. It utilizes the finite-order neighborhood information of nodes in the network topology to learn low-dimensional vector representations of nodes to train the model. This eliminates the need for cumbersome feature engineering due to data changes and results in higher model prediction accuracy. Based on the expanded user social relationship network from link prediction, a random walk algorithm is used to calculate the neighborhood network of each node; initial embedding vector representations of each node are generated based on the neighborhood network structure information; deep embedding representations of nodes are obtained through convolutional layers; and then fully connected layers are used to predict the influence of each node in the user social relationship network. The overall computational framework is as follows: Figure 4 As shown.

[0056] It is understandable that the node importance recognition algorithm based on graph convolutional neural networks is a convolutional neural network (CNN), which is a tool algorithm for discovering important nodes in social relationship networks.

[0057] The discovery of key nodes in the network as users for personalized content generation includes: generating a feature matrix for each social relationship network node based on the number of node hops according to neighborhood relationships, and using a convolutional neural network to train and predict node influence.

[0058] It is understood that the method of using the finite-order neighborhood information of nodes in the network topology to learn the low-dimensional vector representation of nodes to train the model includes: finding L neighboring nodes for each node, selecting the nodes with the closest hop count and the largest degree, and calculating the feature matrix after hop count transformation using the adjacency matrix. The random walk algorithm is one way to construct the neighborhood network in a node importance recognition algorithm based on graph convolutional neural networks.

[0059] Calculate the neighborhood network of each node in the network, that is, calculate the feature matrix of each node in the network.

[0060] The generation of initial embedding vector representations for each node based on neighborhood network structure information includes: calculating the feature matrix of each node in the network using a random walk method based on finite-order neighborhood information in the node network structure.

[0061] Predicting the influence of each node in a user's social relationship network using fully connected layers involves mapping calculations of fully connected layers in a convolutional neural network.

[0062] S4, personalized text content generation incorporating key user attribute information. To accurately generate text content tailored to each user's specific characteristics within a given topic context, the efficient generation capabilities of a large-scale dialogue language model are required. This involves generating corresponding comments or short dialogue texts based on high-value target users identified from social networks and their statements. The generated text content must not only satisfy the contextual relevance and logical consistency within the user's situation but also reflect the positive or negative sentiment expressed in the text.

[0063] Examples of large-scale dialogue language models include ChatGPT, ChatGLM, and Baichuan.

[0064] The entire training process architecture of this invention is shown below. Figure 5 As shown, the input data includes dialogue content, relevant attribute information of identified key users, and positive / negative sentiment information. A personalized text generation model processes this input data to generate several text contents with controllable sentiment that conform to the user's context. The relevant attributes of the identified key users include the user's friend relationships and social relationships with users who have interacted with the user. The positive / negative sentiment information includes two categories: "positive" and "negative" text sentiment tendencies. The personalized text generation model uses an open-source large-scale dialogue language model as its base model. The autoencoder, neighborhood interaction encoder, and sentiment attention encoder in the personalized text generation model encode the dialogue content, key user social relationship attributes, and positive / negative sentiment information respectively, obtaining corresponding encoded feature vectors. Based on the feature vector fusion module in the personalized text generation model, the encoded feature vectors output by each encoder are fused to obtain a fused feature vector. The decoder in the personalized text generation model decodes the fused feature vector. This invention, after identifying target users with high social value, can generate positive / negative sentiment text content based on the target user's and their interacting friends' social relationships and a large-scale dialogue language model, achieving personalized expression within a topic context.

[0065] For example, processing input data through a personalized text generation model includes: a data processing flow, including social relationship data preprocessing and various text data preprocessing, such as extracting social relationship networks from users' social interaction relationships, text data encoding, and positive / negative emotion encoding.

[0066] The personalized text generation model utilizes an autoencoder, a neighborhood interaction encoder, and an emotion attention encoder to encode dialogue content, important user social relationship attributes, and positive / negative emotional information, respectively. This includes: the autoencoder encodes the input text into low-dimensional features via a neural network, i.e., text self-encoding; the graph attention neural network obtains the representation of each node in the neighborhood subgraph through an attention mechanism, and then obtains the interaction features of pairwise node pairs through element-wise vector processing, i.e., neighborhood interaction encoding; and the emotion attention encoder uses multi-head attention and pointwise convolutional transformations to encode the hidden states and semantic interactions between the emotional target word and context words.

[0067] The feature vector fusion module in the personalized text generation model fuses the encoded feature vectors output by each encoder to obtain a fused feature vector. This includes fusing text autoencoding, neighborhood interaction encoding, and sentiment attention encoding through vector concatenation in feature fusion.

[0068] The decoder in the personalized text generation model is used to decode the fused feature vector, including: autoregressive decoding of the fused features through a decoder with a one-way attention mechanism to generate text prediction.

[0069] The reasoning process is as follows Figure 6 As shown, a neighborhood interaction encoder that no longer requires information about the user's social relationships with their friends is needed. During inference, based on the input dialogue content and positive / negative sentiment information, the code passes through the corresponding encoder, feature fusion module, and decoder to generate several candidate text contents with controllable sentiment that conform to the user's context.

[0070] As can be seen from the above embodiments, the present invention can be used to generate personalized text content that is relevant to topics and conforms to the user's context on social media platforms.

[0071] Example 2: This invention provides a personalized topic text content generation system, which includes:

[0072] The user social relationship network construction module is used to obtain topic-related data from social media platforms within the current time period and store the topic-related data in the information database; then, through complex network analysis methods, it uses user interaction data and friend and following lists to construct a social relationship network with users as nodes.

[0073] The graph link prediction module for potential relationships in user social networks is used to predict the most likely pair of nodes to have edge relationships in the user social relationship network that has been constructed under the context of a topic.

[0074] The high social value target discovery module uses a node importance recognition algorithm based on graph convolutional neural networks to discover key nodes in the network as users for personalized content generation. It learns low-dimensional vector representations of nodes by utilizing finite-order neighborhood information in the network topology to train the model. Based on link prediction, it expands the user's social relationship network and uses a random walk algorithm to calculate the neighborhood network of each node in the user's social relationship network. It generates initial embedding vector representations of each node based on neighborhood network structure information. It obtains deep embedding representations of nodes through convolutional layers and then uses fully connected layers to predict the influence of each node in the user's social relationship network.

[0075] The personalized text content generation module, which incorporates important user attribute information, is used to generate sentiment-positive and negative text content based on the social relationships between the target user and their interacting friends, as well as a large-scale dialogue language model, after identifying target users with high social value.

[0076] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0077] The information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0078] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments.

[0079] This invention also provides a computer device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor executes the computer program to implement the steps in any of the above method embodiments.

[0080] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps described in the various method embodiments above.

[0081] This invention also provides an information data processing terminal, which, when executed on an electronic device, provides a user input interface to implement the steps described in the above method embodiments. The information data processing terminal is not limited to mobile phones, computers, or switches.

[0082] This invention also provides a server that, when executed on an electronic device, provides a user input interface to implement the steps described in the above method embodiments.

[0083] This invention provides a computer program product that, when run on an electronic device, enables the electronic device to implement the steps described in the various method embodiments.

[0084] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0085] This invention innovatively integrates the discovery of high-value users within a topic context into the field of text generation, building upon existing large-scale dialogue language models. By identifying important users from the interactions between users within a topic, and incorporating users' neighborhood social attributes and emotional positivity / negativity into the neural network's encoding process through an attention mechanism, it guides text generation. Text generated using this personalized topic-based text generation model improves the fluency of user-contextual dialogue within a topic context, increases overall text topic relevance by 10%, and enhances the accuracy of controllable positive / negative sentiment by 15%, significantly improving the automated generation capability of public opinion guidance texts.

[0086] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for generating personalized topic-based text materials, characterized in that, This method incorporates the discovery of influential users within a given topic into personalized text content generation. It employs graph link prediction and node importance graph neural network methods to discover important users, automatically identifying high-value and influential users from the sparse social relationship network of users in the topic context, and using them as the target audience for personalized text content generation. Simultaneously, it combines a large-scale dialogue language model based on deep learning, integrating feature encoding representations of users' social importance and the positive / negative sentiment of the text, to automatically generate sentiment-controlled personalized text based on users' social characteristics. Specifically, the following steps are included: S1, Building User Social Relationship Networks; S2, predicting potential relationships in a user's social network based on graph links; S3, Discovering High Social Value Targets; S4 generates personalized text content by combining important user attribute information; In step S2, predicting potential relationships in a user's social network based on graph links includes: using the PLNLP algorithm to predict links; based on the constructed user social network, using a negative sample sampling strategy to sample positive and negative samples, where positive samples are node pairs with existing connections in the network, and negative samples are node pairs without existing connections; and then encoding the neighborhood subgraphs of the node pairs through a graph neural network encoder to extract useful relationship information features from the network. The node pair encoding represents the probability of two nodes being connected by an edge calculated using the link scoring function in the link predictor; through the above calculation process, a batch of node pairs most likely to have edge relationships in the user social relationship network already constructed under the topic background are predicted; The personalized text content generation in step S4, which incorporates important user attribute information, includes: The dialogue content, relevant attribute information of the identified important users, and positive and negative sentiment information are used as input data. The input data is processed by a personalized text generation model to generate several emotionally controllable text contents that conform to the user's context. The autoencoder, neighborhood interaction encoder, and sentiment attention encoder in the personalized text generation model are used to encode the dialogue content, social relationship attributes of important users, and positive and negative sentiment information, respectively, to obtain the corresponding encoded feature vectors. Based on the feature vector fusion module in the personalized text generation model, the encoded feature vectors output by each encoder are fused to obtain a fused feature vector; the decoder in the personalized text generation model is used to decode the fused feature vector; after discovering target users with high social value, positive and negative sentiment text content is generated based on the social relationship between the target user and their interactive friends and a large-scale dialogue language model. By combining the neighborhood interaction encoder that the model has already learned about users and their interaction relationships, during inference, only the input dialogue content and the positive and negative emotional information need to be encoded. The feature fusion module integrates the trained neighborhood interaction encoder and, after passing through the model decoder, generates several candidate text contents with controllable emotions that conform to the user's context.

2. The personalized topic text content generation method according to claim 1, characterized in that, In step S1, the construction of the user's social relationship network includes: Data related to current events and topics is obtained from social media platforms, including post and comment text data, user likes, comments, and reposts interaction data, and basic information about users' friend lists and following lists. This data is stored in a database. Then, using complex network analysis methods, a social relationship network with users as nodes is constructed based on user interaction data and friend / following lists.

3. The personalized topic text content generation method according to claim 2, characterized in that, Data related to topics and events within the current time period is obtained from social media platforms. This data is updated regularly, including topic text data and user interaction relationship data. The new data is then integrated into the user's social relationship network.

4. The personalized topic text content generation method according to claim 1, characterized in that, The parameter learning of the entire link prediction model is achieved through the objective function of Pairwise Ranking, which specifically includes: The link prediction model framework includes a neighborhood encoder, a link predictor, negative sampling, and a pairwise ranking objective function. The goal of link prediction is to rank positive samples higher than negative samples. The pairwise ranking objective function uses the idea of ​​pairwise learning to optimize the parameters of the link prediction model, which is consistent with the overall goal of link prediction.

5. The personalized topic text content generation method according to claim 1, characterized in that, In step S3, the discovery of high social value targets includes: using a node importance recognition algorithm based on graph convolutional neural networks to discover key nodes in the network as users for personalized content generation; using finite-order neighborhood information of nodes in the network topology to learn low-dimensional vector representations of nodes to train the model; predicting the expanded user social relationship network based on links; using a random walk algorithm to calculate the neighborhood network of each node in the user social relationship network; generating initial embedding vector representations of each node based on neighborhood network structure information; obtaining deep embedding representations of nodes through convolutional layers; and then using fully connected layers to predict the influence of each node in the user social relationship network.

6. The personalized topic text content generation method according to claim 1, characterized in that, Among the generated text content with controllable emotions that conform to the user's context, the important user-related attributes that have been discovered include the user's friend relationships and social relationships with users who have interacted with the user. The positive and negative emotional information includes two types of text emotional tendencies: positive and negative.

7. A personalized topic text and public opinion material generation system, characterized in that, The personalized topic text content generation method according to any one of claims 1 to 6, the system comprising: The user social relationship network construction module is used to obtain relevant data on topics and events within the current time period from social media platforms and store them in the information database; then, through complex network analysis methods, it uses user interaction data and friend and following lists to construct a social relationship network with users as nodes. The graph link prediction module for potential relationships in user social networks is used to predict the most likely pair of nodes to have edge relationships in the user social relationship network that has been constructed under the context of a topic. The high social value target discovery module uses a node importance recognition algorithm based on graph convolutional neural networks to discover key nodes in the network as users for personalized content generation. It learns low-dimensional vector representations of nodes by utilizing finite-order neighborhood information in the network topology to train the model. Based on link prediction, it expands the user's social relationship network and uses a random walk algorithm to calculate the neighborhood network of each node in the user's social relationship network. It generates initial embedding vector representations of each node based on the neighborhood network structure information. It obtains deep embedding representations of nodes through convolutional layers and then uses fully connected layers to predict the influence of each node in the user's social relationship network. The personalized text content generation module, which incorporates important user attribute information, is used to generate sentiment-positive and negative text content based on the social relationships between the target user and their interacting friends, as well as a large-scale dialogue language model, after identifying target users with high social value.

Images