Method, device and storage medium for determining message category

Abstract

An embodiment of the present specification provides a method, device and storage medium for determining a message category, the method comprising: obtaining a plurality of annotated information propagation trees; wherein each information propagation tree comprises a source message and a plurality of forwarded messages related to the source message; the forwarded message is annotated with a first category of the forwarded message; modeling the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message; fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; determining the first category of each forwarded message in the information propagation tree to be classified according to the first classification condition, thereby improving the accuracy of message category determination.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of artificial intelligence technology, and more particularly to a method, device, and storage medium for determining a message category. Background Art

[0002] The attention mechanism originates from research on human vision. In cognitive science, due to information processing bottlenecks, humans selectively focus on a subset of all information while ignoring other visible information. This mechanism is often referred to as the attention mechanism. In recent years, the attention mechanism has been widely applied to various natural language processing tasks based on deep learning. Thanks to the attention mechanism, existing work can better capture and identify rumor-related messages and users, while also providing a certain degree of interpretability.

[0003] In the above-mentioned method of identifying whether a message is a rumor, locating the messages and words that affect rumor discrimination can capture data deviations of different authenticity to a certain extent. However, due to the low repetition rate of news events, such methods cannot perceive messages and words that do not appear in the training corpus, and their generalization ability is poor. Locating users who have a key influence on the spread of rumors can capture the characteristics of malicious users to a certain extent, but such methods ignore the explicit textual meaning when rumors spread.

[0004] Therefore, the existing technology still has the problem of low accuracy in identifying whether a message is a rumor. Summary of the Invention

[0005] The purpose of the embodiments of this specification is to provide a method, device, and storage medium for determining a message category to improve the accuracy of determining a message category.

[0006] To solve the above problem, an embodiment of the present specification provides a method for determining a message category, the method comprising: obtaining a plurality of annotated information propagation trees; wherein each information propagation tree comprises a source message and a plurality of forwarded messages related to the source message; the forwarded message is annotated with the first category of the forwarded message; the propagation direction and interaction characteristics of the forwarded information are modeled to extract the first feature of each forwarded message; the first feature of each forwarded message and the first category corresponding to each forwarded message are fitted according to a first loss function to obtain a first classification condition for the forwarded message; and the first category of each forwarded message in the information propagation tree to be classified is determined according to the first classification condition.

[0007] Furthermore, the source message is marked with the second category of the source message, and the method also includes: assigning different weights to forwarded messages of different categories according to preset rules; performing weighted summation of each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain the second feature of the source message in each information propagation tree; fitting the second feature of the source message in each information propagation tree and the second category of each source message according to a second loss function to obtain the second classification condition of the source message; determining the second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

[0008] Furthermore, assigning different weights to different categories of forwarded messages according to preset rules includes:

[0009]

[0010] a i represents the weight of the forwarded message corresponding to the i-th node, Represents the first category of the corresponding forwarding message of the node.

[0011] Furthermore, the modeling of the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message includes: constructing each information propagation tree into an undirected graph structure consisting of nodes and edges; wherein the nodes represent the forwarded messages and the edges represent the propagation relationships of the forwarded messages; splitting the undirected graph structure into a first directed graph structure and a second directed graph structure according to the propagation direction and interaction characteristics of the forwarded information; the first directed graph structure represents the information flow from the parent node to the child node, and the second directed graph structure represents the information flow from the child node to the parent node; using a graph recurrent neural network to perform feature fusion on adjacent nodes in the first directed graph structure and the second directed graph structure respectively to extract the first feature of each forwarded message.

[0012] Furthermore, the use of a graph recurrent neural network to perform feature fusion on adjacent nodes in the first directed graph structure and the second directed graph structure respectively to extract the first feature of each forwarded message includes: adding a self-loop to each leaf node in the first directed graph structure and the second directed graph structure respectively; normalizing the first directed graph structure and the second directed graph structure after adding the self-loop; and using a graph recurrent neural network to perform feature fusion on adjacent nodes in the normalized first directed graph structure and the normalized second directed graph structure respectively to extract the first feature of each forwarded message.

[0013] Furthermore, the normalizing operation on the first directed graph structure and the second directed graph structure after adding the self-loop includes:

[0014]

[0015]

[0016] in, Indicates A p ,A c A diagonal matrix with diagonal elements equal to the node degrees, p ,A c Represent the adjacency matrix of the first directed graph structure and the adjacency matrix of the second directed graph structure respectively, represents the normalized adjacency matrix.

[0017] To solve the above problems, an embodiment of the present specification also provides a message category determination device, which includes: an acquisition module for acquiring multiple annotated information propagation trees; wherein each information propagation tree includes a source message and multiple forwarded messages related to the source message; the forwarded message is annotated with the first category of the forwarded message; an extraction module for modeling the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message; a first fitting module for fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; a first determination module for determining the first category of each forwarded message in the information propagation tree to be classified according to the first classification condition.

[0018] Furthermore, the source message is marked with the second category of the source message, and the device also includes: a weighting module, used to assign different weights to forwarded messages of different categories according to preset rules; a summation module, used to perform weighted summation of each forwarded message in the information propagation tree according to the weight of the forwarded message, and obtain the second feature of the source message in each information propagation tree; a second fitting module, used to fit the second feature of the source message in each information propagation tree and the second category of each source message according to a second loss function, and obtain the second classification condition of the source message; a second determination module, used to determine the second category of the source message in the information propagation tree to be classified based on the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

[0019] To solve the above problems, an embodiment of the present specification also provides an electronic device, comprising: a memory for storing a computer program; a processor for executing the computer program to implement: obtaining multiple information propagation trees with annotations; wherein each information propagation tree includes a source message and multiple forwarding messages related to the source message; the forwarding message is annotated with the first category of the forwarding message; modeling the propagation direction and interaction characteristics of the forwarding information to extract the first feature of each forwarding message; fitting the first feature of each forwarding message and the first category corresponding to each forwarding message according to a first loss function to obtain a first classification condition for the forwarding message; determining the first category of each forwarding message in the information propagation tree to be classified according to the first classification condition.

[0020] To solve the above problems, an embodiment of the present specification also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed, implement: obtaining multiple annotated information propagation trees; wherein each information propagation tree includes a source message and multiple forwarding messages related to the source message; the forwarding message is annotated with the first category of the forwarding message; the propagation direction and interaction characteristics of the forwarding information are modeled to extract the first feature of each forwarding message; the first feature of each forwarding message and the first category corresponding to each forwarding message are fitted according to a first loss function to obtain a first classification condition for the forwarding message; and the first category of each forwarding message in the information propagation tree to be classified is determined according to the first classification condition.

[0021] It can be seen from the technical solutions provided by the above embodiments of this specification that, in the embodiments of this specification, multiple annotated information propagation trees can be obtained; wherein, each information propagation tree includes a source message and multiple forwarding messages related to the source message; the forwarding message is annotated with the first category of the forwarding message; the propagation direction and interaction characteristics of the forwarding information are modeled to extract the first feature of each forwarding message; the first feature of each forwarding message and the first category corresponding to each forwarding message are fitted according to the first loss function to obtain the first classification condition of the forwarding message; the first category of each forwarding message in the information propagation tree to be classified is determined according to the first classification condition. The method provided in the embodiments of this specification improves the accuracy of forwarding message category identification by modeling the propagation direction and interaction characteristics of the forwarding message in the information propagation tree and using them for forwarding message category identification; further, by extracting the characteristics of each forwarding message in the information propagation tree to identify the category of the source message in the information propagation tree, a pre-emptive and fine-grained message-level task is constructed, thereby improving the accuracy of source message category identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 This is a flow chart of a method for determining a message category according to an embodiment of this specification;

[0024] Figure 2 This is a flow chart of another method for determining a message category according to an embodiment of this specification;

[0025] Figure 3 This is a functional structure diagram of an electronic device according to an embodiment of this specification;

[0026] Figure 4 This is a functional structure diagram of a device for determining a message category according to an embodiment of this specification. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of this specification to clearly and completely describe the technical solutions in the embodiments of this specification. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this specification.

[0028] Early rumor detection models were primarily based on feature engineering, with researchers extracting features based on factors such as the content being spread, user credibility, and influence. In recent years, deep learning has provided a more accessible solution, avoiding the tedious feature extraction process and subjective cognitive biases inherent in feature engineering. Information propagation trees are organized into different forms (time series, tree structures, or graphs), allowing them to be modeled using existing deep learning models to obtain a global representation of the propagation tree. Researchers have also leveraged stance information to aid rumor detection, using a multi-task learning framework to improve data utilization and mitigate overfitting. Other researchers have used hierarchical attention and co-attention mechanisms to capture sentences, words, and users that aid in rumor detection.

[0029] However, in the above-mentioned method of identifying whether a message is a rumor, locating the messages and words that affect rumor discrimination can capture data deviations of rumors of different authenticity to a certain extent. However, due to the low repetition rate of news events, such methods cannot perceive messages and words that do not appear in the training corpus, and their generalization ability is poor; locating users who have a key influence on the spread of rumors can capture the characteristics of malicious users to a certain extent, but such methods ignore the explicit textual meaning when rumors spread.

[0030] After the applicant explored the rumor propagation mechanism from a quantitative perspective, it was found that there are three points that are of enlightenment significance to rumor governance: (1) The reason why rumors spread faster and wider than ordinary messages may be that rumors bring more novel information (information novelty); (2) The echo chamber member effect is obvious in the process of rumor propagation, that is, the distorted and biased opinions formed by a small number of users are easily spread and discussed on social networks, and are regarded as the truth by most users who do not know the truth, thus forming a viral rumor propagation; (3) There are three types of roles in the spread of rumors: seeds firmly believe in the authenticity of the message, agnostics question the authenticity of the message, and others hold other irrelevant opinions. Considering that if a more fine-grained message-level task can be constructed as a prerequisite for the rumor discrimination task, and the propagation direction and interaction characteristics of the more fine-grained message can be modeled, it is expected to solve the problem of inaccurate rumor discrimination in the existing technology. Based on this, the embodiments of this specification provide a message category determination method, device and storage medium to improve the accuracy of message category determination.

[0031] See also Figure 1 . An embodiment of this specification provides a method for determining a message category. In an embodiment of this specification, the subject that executes the message category determination method may be an electronic device with a logical operation function, and the electronic device may be a server. The server may be an electronic device with a certain computing and processing capability. It may have a network communication unit, a processor, a memory, etc. Of course, the server is not limited to the above-mentioned electronic device with a certain entity, and it may also be software running in the above-mentioned electronic device. The server may also be a distributed server, and may be a system with multiple processors, memories, network communication modules, etc. operating in collaboration. Alternatively, the server may also be a server cluster formed by several servers. The method may include the following steps.

[0032] S110: Acquire multiple information propagation trees with annotations; wherein each information propagation tree includes a source message and multiple forwarding messages related to the source message; the forwarding messages are annotated with a first category of the forwarding messages.

[0033] In the embodiments of this specification, information dissemination is typically performed in a tree format. A source message may be the starting point of the information dissemination tree, and forwarded messages may be branches of the tree. Forwarded messages may include comments on the source message, comments on other forwarded messages, and so on. Each information dissemination tree may include a source message and multiple forwarded messages related to the source message. For example, a source message may be an article, a Weibo post, a tweet, etc., and forwarded messages may be comments on these articles, Weibo posts, tweets, etc., or comments on comments, etc.

[0034] Rumor identification is a classification task performed at the level of the propagation tree. Given an unverified source message r0 and the corresponding forwarded messages {r1, r2, ... r T}, predicting the authenticity of the propagation tree and classifying the source message (second category) as true, false, and unverified. Information tipping point identification is a message-level classification task. For each forwarded message involved in the propagation tree, it is classified (first category) as amplifying, denying, clarifying, or null. Information tipping points represent forwarded messages that promote rumor propagation or provide indicative information for rumor identification.

[0035] In the embodiments of this specification, the annotated information dissemination tree may be an information dissemination tree that annotates source messages and / or forwarded messages. Specifically, the forwarded messages in each information dissemination tree may be manually annotated in advance to indicate the first category of the forwarded message, i.e., whether the forwarded message belongs to the topic-opening category, the objection-questioning category, the clarification category, or the meaningless category. Similarly, the source messages in each information dissemination tree may also be manually annotated in advance to indicate the second category of the source message, i.e., whether the source message belongs to the true category, the false category, or the unverified category.

[0036] In the embodiment of the present specification, a plurality of information propagation trees with labels may be obtained, and in each information propagation tree, each forwarded message is labeled with the category of the forwarded message.

[0037] S120: Modeling the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message.

[0038] Content coherence can be presented by the text similarity with the previous and next texts. The similarity with the previous text is determined by the ratio of the number of identical words between the message and the parent node to the number of words in the longest text; the similarity with the next text is determined by calculating the similarity with each child node text and taking the average.

[0039] The average values of the two similarities in different events were calculated for the four types of forwarded messages. The results showed that: (1) Overall, for data from different events, the forwarded messages of various categories showed a high degree of consistency in text similarity, indicating that the information trigger points also had similar text interaction characteristics in different event contexts. (2) For forwarded messages of different categories, the forwarded messages in the topic-starting category had a low similarity with the parent node, but a very high similarity with the child nodes, indicating that the child nodes mostly started discussions based on new topics; the forwarded messages in the objection-questioning category had a low similarity with both the parent node and the child node, and mostly contained emotional words with negative attitudes; the forwarded messages in the clarification category had a medium similarity with the parent node and the child node, and were more likely to be elaborated based on the previous text.

[0040] By modeling the propagation direction and interaction characteristics of forwarded information, the first feature of each forwarded message can be extracted. The first feature can be a feature vector, and the first feature of the forwarded message is extracted in the form of a vector.

[0041] Graph neural networks have garnered increasing attention in recent years. Information propagation trees can also be modeled using graph neural networks, where nodes represent individual messages and edges represent forwarding relationships between messages. However, current graph neural networks are primarily used to model homogeneous node graphs, where the attributes of connected nodes are relatively similar. Therefore, message passing can significantly enrich node features. However, in information propagation trees, nodes connected by forwarding relationships are often heterogeneous, making existing graph neural networks inadequate for modeling text propagation on social networks. Furthermore, rumor detection involves classifying this graph, while current methods for learning representations of the entire graph often employ mean pooling and attention pooling, resulting in a relatively coarse representation.

[0042] Based on this, in some embodiments, based on existing graph neural networks, node features can be updated by considering the propagation direction and interaction characteristics of messages, and the entire graph can be represented using the tipping point information. Specifically, the propagation direction and interaction characteristics of forwarded messages are modeled to extract the first feature of each forwarded message, including:

[0043] Each information propagation tree is constructed as an undirected graph structure consisting of nodes and edges; wherein the nodes represent forwarded messages and the edges represent the propagation relationship of the forwarded messages; the undirected graph structure is split into a first directed graph structure and a second directed graph structure according to the propagation direction and interaction characteristics of the forwarded information; the first directed graph structure represents the information flow from the parent node to the child node, and the second directed graph structure represents the information flow from the child node to the parent node; a graph recurrent neural network is used to perform feature fusion on adjacent nodes in the first directed graph structure and the second directed graph structure respectively to extract the first feature of each forwarded message.

[0044] Specifically, each information propagation tree can be constructed as a graph consisting of nodes and edges. in Represents a set of nodes, where a node represents a message, and ε represents a set of edges, where edges are generated by the forwarding relationship between forwarded messages. The initial representation of each node is obtained by the pre-trained language model BERT, and the representation of the [CLS] flag in the sentence is taken as the sentence representation. In the overall training of the model, the initialization sentence representation is fixed to speed up the training. Stacked together, we can get the node feature matrix of the graph

[0045] Typically, the edges in an undirected graph are represented by a symmetric adjacency matrix To express it, if there is an edge between node i and node j, then A ij =A ji =1. However, applying such traditional graph structure representation to propagation graph modeling ignores the modeling of information flow direction and information interaction. Therefore, in the embodiment of this specification, the adjacency matrix of the undirected graph structure can be split to form two directed graph structure representations A p ,A c , respectively representing the information flow from parent nodes to child nodes and the information flow from child nodes to parent nodes.

[0046] In some embodiments, the using a graph recurrent neural network to perform feature fusion on adjacent nodes in the first directed graph structure and the second directed graph structure to extract a first feature of each forwarded message includes:

[0047] A self-loop is added to each leaf node in the first directed graph structure and the second directed graph structure respectively; the first directed graph structure and the second directed graph structure after adding the self-loop are normalized; and a graph recurrent neural network is used to perform feature fusion on adjacent nodes in the normalized first directed graph structure and the normalized second directed graph structure respectively to extract the first feature of each forwarded message.

[0048] Specifically, the normalizing operation on the first directed graph structure and the second directed graph structure after adding the self-loop includes:

[0049]

[0050]

[0051] in, Indicates A p ,A c A diagonal matrix with diagonal elements equal to the node degrees, Represents a node set, A p ,A c Represent the adjacency matrix of the first directed graph structure and the adjacency matrix of the second directed graph structure respectively, represents the normalized adjacency matrix.

[0052] Graph recurrent neural networks use recurrent neural units as a way to pass messages, and at the same time combine the efficient feature fusion method of graph convolutional neural networks, which can be expanded into an L-layer structure. When updating the node representation of the l-th layer, the node representation of the previous layer, i.e., the l-1 layer, is required. Specifically, taking the direction from the parent node to the child node as an example, the adjacency matrix is first used to fuse the features of the connected nodes of the node, as shown in formula (3), to obtain a temporary aggregate representation. Then the node of the previous layer is represented The input is a temporary aggregate representation, the hidden state is initialized, and message passing is performed using the calculation method of the long-short-term memory unit (LSTM) to obtain the node representation of the lth layer. In the embodiment of this specification, a one-layer graph recurrent neural network is used as an example.

[0053]

[0054] S130: Fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to the first loss function to obtain a first classification condition for the forwarded message.

[0055] The first loss function can characterize the error between the first category (predicted value) of the fitted forwarded message and the first category (annotated value) of the annotated forwarded message. The loss function maps the value of a random event or its related random variable to a non-negative real number to represent the "risk" or "loss" of the random event. In applications, loss functions are often used as learning criteria in conjunction with optimization problems, i.e., models are solved and evaluated by minimizing the loss function.

[0056] In some embodiments, a feed forward neural network (FFN) may be used to identify the first category to which the forwarded message belongs, and a cross-entropy loss function may be used for fitting:

[0057]

[0058]

[0059] in, represents the first category of the fitted (predicted) forwarded message, It represents the feature obtained by using a one-layer graph recurrent neural network (the first feature of the forwarded message). represents the first loss function, represents the fitted value, Indicates the marked value, L t Indicates the number of fits.

[0060] In some embodiments, the first classification condition may be a mapping relationship between the first feature of each forwarded message and the first category corresponding to each forwarded message, for example, f(x1)=y′1, where x1 represents the first feature of the forwarded message, and y′1 represents the fitted (predicted) first category of the forwarded message.

[0061] S140: Determine a first category of each forwarded message in the information propagation tree to be classified according to the first classification condition.

[0062] After obtaining the first classification condition for the forwarded message, the first category of each forwarded message in the information propagation tree to be classified can be determined based on the first classification condition. Specifically, the information propagation tree to be classified can be obtained, and the first features of each forwarded message in the information propagation tree to be classified can be extracted according to S120. These first features are used as input to determine the first category of each forwarded message based on the first classification condition.

[0063] The above method for determining the first category of each forwarded message in the information propagation tree to be classified improves the accuracy of forwarded message category identification by modeling the propagation direction and interaction characteristics of the forwarded message in the information propagation tree and using them for forwarded message category identification.

[0064] In some embodiments, to further identify the second category of the source message, such as Figure 2 As shown, the method further includes the following steps.

[0065] S210: Assign different weights to different categories of forwarded messages according to preset rules.

[0066] Through research and investigation, it was found that the proportion of forwarded messages that oppose and question in the false category of source messages is much higher than that of the other two types of source messages; the proportion of clarification statements in the false category of source messages is also relatively high, and they appear frequently over a long period of time in the early stage. Therefore, identifying different categories of forwarded messages will also help to determine the final source message.

[0067] In some embodiments, different weights are assigned to different categories of forwarded messages according to preset rules, so that the forwarded messages marked as meaningless (null) have as little impact on the weight of the source message as possible, and this result is used as the weight. Specifically, assigning different weights to different categories of forwarded messages according to preset rules includes:

[0068]

[0069] a i represents the weight of the forwarded message corresponding to the i-th node, Indicates the first category of the corresponding forwarding message of the i-th node.

[0070] S220: performing weighted summation on each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain a second feature of the source message in each information propagation tree.

[0071] Specifically, performing weighted summation on each forwarded message in the information propagation tree includes:

[0072]

[0073] Wherein, c is a matrix representing the second feature of the source message.

[0074] S230: Fitting the second feature of the source message in each information propagation tree and the second category of each source message according to the second loss function to obtain a second classification condition for the source message.

[0075] In some embodiments, a feed forward neural network may be used to identify the second category to which the source message belongs, and a cross-entropy loss function may be used for fitting:

[0076]

[0077]

[0078] Among them, y v The second category of source messages representing the fit (prediction), Represents the second loss function, y v,j represents the fitted value, Indicates the marked value, Lv Indicates the number of fits.

[0079] In some embodiments, the second classification condition may be a mapping relationship between the second feature of each source message and the second category corresponding to each source message, for example, f(x2)=y′2, where x2 represents the second feature of the source message, and y′2 represents the second category of the fitted (predicted) source message.

[0080] S240: Determine a second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

[0081] After obtaining the second classification condition for the source message, the second category of the source message in the information propagation tree to be classified can be determined based on the second classification condition. Specifically, the information propagation tree to be classified can be obtained, and the first features of each forwarded message in the information propagation tree to be classified can be extracted according to S120. These first features are used as input, and the first category of each forwarded message is determined based on the first classification condition. Then, according to S210 and S220, a weighted sum is performed on each forwarded message in the information propagation tree to be classified to obtain the second feature of the source message in the information propagation tree to be classified. The second feature is used as input, and the second category of the source message is determined based on the second classification condition.

[0082] The method provided in the embodiments of this specification improves the accuracy of forwarding message category identification by modeling the propagation direction and interaction characteristics of forwarded messages in the information propagation tree and using them to identify the categories of forwarded messages; further, by extracting the characteristics of each forwarded message in the information propagation tree to identify the category of the source message in the information propagation tree, a pre-emptive and fine-grained message-level task is constructed, thereby improving the accuracy of source message category identification.

[0083] The method embodiments provided in the above embodiments of this specification can be executed in a mobile terminal, a computer terminal or similar electronic devices. Figure 3 This is a functional structure diagram of an electronic device according to an embodiment of this specification. The electronic device may include a memory and a processor.

[0084] In some embodiments, the memory can be used to store the computer program and / or module, and the processor implements various functions of the shared device location prediction method by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function; the data storage area can store data created according to the use of the user terminal. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0085] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The processor may execute the computer instructions to implement the following steps: obtaining multiple annotated information propagation trees; each information propagation tree includes a source message and multiple forwarded messages related to the source message; the forwarded messages are annotated with a first category of the forwarded messages; modeling the propagation direction and interaction characteristics of the forwarded messages to extract a first feature of each forwarded message; fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; and determining a first category for each forwarded message in the information propagation tree to be classified based on the first classification condition.

[0086] In some embodiments, the processor can also execute the computer instructions to implement the following steps: assigning different weights to forwarded messages of different categories according to preset rules; performing weighted summation of each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain the second feature of the source message in each information propagation tree; fitting the second feature of the source message in each information propagation tree and the second category of each source message according to the second loss function to obtain the second classification condition of the source message; determining the second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

[0087] In the embodiments of this specification, the functions and effects specifically achieved by the electronic device can be explained in comparison with other embodiments and will not be repeated here.

[0088] Figure 4 This is a functional structural diagram of a device for determining a message category according to an embodiment of the present specification. The device may specifically include the following structural modules.

[0089] The acquisition module 410 is configured to acquire a plurality of annotated information propagation trees, wherein each information propagation tree includes a source message and a plurality of forwarded messages related to the source message; the forwarded messages are annotated with a first category of the forwarded messages;

[0090] An extraction module 420 is configured to model the propagation direction and interaction characteristics of the forwarded information to extract a first feature of each forwarded message;

[0091] A first fitting module 430 is configured to fit the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message;

[0092] The first determining module 440 is configured to determine a first category of each forwarded message in the information propagation tree to be classified according to the first classification condition.

[0093] In some implementations, the apparatus may further include:

[0094] The weighting module is used to assign different weights to different categories of forwarded messages according to preset rules;

[0095] a summing module, configured to perform weighted summation on each forwarded message in the information propagation tree according to the weight of the forwarded message, to obtain a second feature of the source message in each information propagation tree;

[0096] A second fitting module is used to fit the second feature of the source message in each information propagation tree and the second category of each source message according to the second loss function to obtain a second classification condition of the source message;

[0097] The second determining module is configured to determine a second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

[0098] An embodiment of the present specification also provides a computer-readable storage medium for a message category determination method, wherein the computer-readable storage medium stores computer program instructions, which, when executed, implement: obtaining a plurality of annotated information propagation trees; wherein each information propagation tree includes a source message and a plurality of forwarded messages related to the source message; the forwarded message is annotated with the first category of the forwarded message; modeling the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message; fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; and determining the first category of each forwarded message in the information propagation tree to be classified according to the first classification condition.

[0099] In some embodiments, when the computer program instructions are executed, the following are also achieved: assigning different weights to forwarded messages of different categories according to preset rules; performing weighted summation of each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain the second feature of the source message in each information propagation tree; fitting the second feature of the source message in each information propagation tree and the second category of each source message according to a second loss function to obtain the second classification condition of the source message; determining the second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

[0100] In the embodiments of this specification, the above-mentioned storage medium includes but is not limited to random access memory (Random Access Memory, RAM), read-only memory (Read-Only Memory, ROM), cache (Cache), hard disk (Hard Disk Drive, HDD) or memory card (Memory Card). The memory can be used to store the computer program and / or module, and the memory can mainly include a storage program area and a storage data area, wherein the storage program area can store an operating system, at least one application required for a function, etc.; the storage data area can store data created according to the use of the user terminal, etc. In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory. In the embodiments of this specification, the functions and effects specifically implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other embodiments and will not be repeated here.

[0101] It should be noted that the various embodiments in this specification are described in a progressive manner. References to the same or similar parts of the various embodiments are sufficient. Each embodiment focuses on the differences from the other embodiments. In particular, since the apparatus and device embodiments are generally similar to the method embodiments, their descriptions are relatively simple. For relevant details, references to the descriptions of the method embodiments are sufficient.

[0102] After reading this specification, those skilled in the art can conceive of any combination of some or all of the embodiments listed in this specification without creative work, and these combinations are also within the scope of disclosure and protection of this specification.

[0103] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0104] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the description of the method embodiments.

[0105] This specification is applicable to numerous general-purpose and special-purpose computer system environments or configurations, such as personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the above.

[0106] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0107] Although the present specification is described through embodiments, those skilled in the art will appreciate that there are many modifications and variations to the present specification without departing from the spirit of the present specification. It is intended that the appended claims include these modifications and variations without departing from the spirit of the present specification.

Claims

1. A method for determining a message category, characterized in that: The method comprises: Acquire a plurality of information propagation trees with annotations, wherein each information propagation tree includes a source message and a plurality of forwarded messages related to the source message; the forwarded messages are annotated with a first category of the forwarded messages; Model the propagation direction and interaction characteristics of forwarded information to extract the first feature of each forwarded message; Fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to the first loss function to obtain a first classification condition for the forwarded message; Determining a first category of each forwarded message in the information propagation tree to be classified according to the first classification condition; The source message is marked with the second category of the source message, and the method further includes: Assign different weights to different categories of forwarded messages according to preset rules; performing weighted summation on each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain a second feature of the source message in each information propagation tree; Fitting the second feature of the source message in each information propagation tree and the second category of each source message according to the second loss function to obtain a second classification condition of the source message; The second category of the source message in the information propagation tree to be classified is determined according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

2. The method according to claim 1, characterized in that Assigning different weights to different categories of forwarded messages according to preset rules includes: a i represents the weight of the forwarded message corresponding to the i-th node, Indicates the first category of the corresponding forwarding message of the i-th node.

3. The method according to claim 1, characterized in that The modeling of the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message includes: Each information propagation tree is constructed as an undirected graph structure consisting of nodes and edges; where nodes represent forwarded messages and edges represent the propagation relationship of forwarded messages; Splitting the undirected graph structure into a first directed graph structure and a second directed graph structure according to the propagation direction and interaction characteristics of the forwarded information; the first directed graph structure represents the flow of information from the parent node to the child node, and the second directed graph structure represents the flow of information from the child node to the parent node; A graph recurrent neural network is used to perform feature fusion on adjacent nodes in the first directed graph structure and the second directed graph structure respectively to extract a first feature of each forwarded message.

4. The method according to claim 3, characterized in that The using a graph recurrent neural network to perform feature fusion on adjacent nodes in the first directed graph structure and the second directed graph structure respectively to extract a first feature of each forwarded message includes: Adding a self-loop to each leaf node in the first directed graph structure and the second directed graph structure respectively; performing a normalization operation on the first directed graph structure and the second directed graph structure after adding the self-loop; A graph recurrent neural network is used to perform feature fusion on adjacent nodes in the normalized first directed graph structure and the normalized second directed graph structure, respectively, to extract a first feature of each forwarded message.

5. The method according to claim 4, characterized in that The normalizing operation on the first directed graph structure and the second directed graph structure after adding the self-loop includes: in, Indicates A p ,A c A diagonal matrix with diagonal elements equal to the node degrees, Represents a node set, A p ,A c Represent the adjacency matrix of the first directed graph structure and the adjacency matrix of the second directed graph structure respectively, represents the normalized adjacency matrix.

6. A message category determination device, characterized in that: The device comprises: An acquisition module, configured to acquire a plurality of annotated information propagation trees, wherein each information propagation tree includes a source message and a plurality of forwarded messages related to the source message; the forwarded messages are annotated with a first category of the forwarded messages; An extraction module, configured to model the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message; A first fitting module, configured to fit the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; A first determining module, configured to determine a first category of each forwarded message in the information propagation tree to be classified according to the first classification condition; The source message is marked with the second category of the source message, and the apparatus further includes: The weighting module is used to assign different weights to different categories of forwarded messages according to preset rules; a summing module, configured to perform weighted summation on each forwarded message in the information propagation tree according to the weight of the forwarded message, to obtain a second feature of the source message in each information propagation tree; A second fitting module is used to fit the second feature of the source message in each information propagation tree and the second category of each source message according to the second loss function to obtain a second classification condition of the source message; The second determining module is configured to determine a second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

7. An electronic device, characterized in that: include: memory for storing computer programs; A processor, configured to execute the computer program to implement: obtaining a plurality of annotated information propagation trees; wherein each information propagation tree includes a source message and a plurality of forwarded messages related to the source message; the forwarded message is annotated with the first category of the forwarded message; the source message is annotated with the second category of the source message; modeling the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message; fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; determining the first category of each forwarded message in the information propagation tree to be classified according to the first classification condition; assigning different weights to forwarded messages of different categories according to a preset rule; performing weighted summation on each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain the second feature of the source message in each information propagation tree; fitting the second feature of the source message in each information propagation tree and the second category of each source message according to a second loss function to obtain a second classification condition for the source message; The second category of the source message in the information propagation tree to be classified is determined according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

8. A computer-readable storage medium, characterized in that Computer instructions are stored thereon, which, when executed, implement the following: obtaining multiple annotated information propagation trees; wherein each information propagation tree includes a source message and multiple forwarding messages related to the source message; the forwarding message is annotated with the first category of the forwarded message, and the source message is annotated with the second category of the source message; modeling the propagation direction and interaction characteristics of the forwarded information to extract the first feature of each forwarded message; fitting the first feature of each forwarded message and the first category corresponding to each forwarded message according to a first loss function to obtain a first classification condition for the forwarded message; determining the first category of each forwarded message in the information propagation tree to be classified according to the first classification condition; assigning different weights to forwarded messages of different categories according to a preset rule; performing weighted summation on each forwarded message in the information propagation tree according to the weight of the forwarded message to obtain the second feature of the source message in each information propagation tree; fitting the second feature of the source message in each information propagation tree and the second category of each source message according to a second loss function to obtain a second classification condition for the source message; determining the second category of the source message in the information propagation tree to be classified according to the second classification condition and the first category of each forwarded message in the information propagation tree to be classified.

Images