CN113449204B - Social event classification method and device based on local aggregation graph attention network - Google Patents
Social event classification method and device based on local aggregation graph attention network Download PDFInfo
- Publication number
- CN113449204B CN113449204B CN202110787860.4A CN202110787860A CN113449204B CN 113449204 B CN113449204 B CN 113449204B CN 202110787860 A CN202110787860 A CN 202110787860A CN 113449204 B CN113449204 B CN 113449204B
- Authority
- CN
- China
- Prior art keywords
- social
- graph
- events
- event
- social events
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/285—Clustering or classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The invention discloses a social event classification method and device based on a local aggregation graph attention network, which can classify social events of a network and a real world and provide decision support for processing the social events: extracting time, place, personnel, organization and keyword information in news text data, modeling to form a heterogeneous information network, extracting a meta-mode of the heterogeneous information network, selecting a plurality of meta-paths with different semantics, calculating the similarity between two social events through the meta-paths, and splicing the semantic features of the social events and the social influence features of the social events to obtain the fusion features of the social events; using the obtained similarity as the edge weight among the nodes of the social event classification graph, using the obtained mixed features as the node features, and constructing the social event classification graph; and constructing and training a social event classification model, inputting the social event classification diagram into the trained social event classification model, and outputting the category of the social event.
Description
Technical Field
The invention belongs to the technical field of network public opinion analysis, data mining and social event analysis, and particularly relates to a social event classification method and device based on a local aggregation graph attention network.
Background
The occurrence of social events such as counseling, fighting and collaboration can have a significant impact on society. The classification of social events from large heterogeneous open source news media is a current research hotspot. In the past researches, the event classification methods mainly regard social events as homoword or homoelement co-occurrence graphs, and the methods achieve certain effects, but the methods are still vague and poor in interpretability when modeling the social events, and the main reason is that the previous researches do not consider different types of elements contained in the events, such as personnel, time, positions and the like. The modeling of the heterogeneous elements and the relation thereof in the social event is important for realizing interpretable and accurate classification, and the invention aims to provide a novel social event classification method to solve the existing problems.
Disclosure of Invention
Aiming at the problems, the invention discloses a social event classification method and device based on a local aggregation graph attention network, which can classify social events of a network and the real world and provide decision support for processing the social events. The technical scheme is as follows: a social event classification method based on a local aggregation graph attention network comprises the following steps:
Step 1: collecting news text data related to social events, and extracting time, place, personnel, organization and keyword information in the news text data;
step 2: modeling the time, place, personnel, organization and keyword information of the extracted social events into a heterogeneous information network, and extracting meta-modes of the heterogeneous information network;
and step 3: selecting a plurality of meta-paths with different semantics according to the meta-mode of the extracted heterogeneous information network, and calculating the similarity between two social events through the meta-paths;
and 4, step 4: obtaining semantic features of the social events from news text data related to the social events; modeling political influence, public opinion influence and emotional polarity of the social events into social influence characteristics of the social events, and splicing semantic characteristics of the social events and the social influence characteristics of the social events to obtain fusion characteristics of the social events;
and 5: using the similarity between the two obtained social events as the edge weight between the nodes of the social event classification graph, using the fusion characteristics of the obtained social events as the node characteristics, and constructing the social event classification graph;
step 6: and constructing and training a social event classification model, wherein the social event classification model comprises a local aggregation graph neural network GraphSAGE layer, a self-supervision graph attention network SuperGAT layer and a logic classification layer which are sequentially arranged, inputting the social event classification graph into the trained social event classification model, and outputting the category of the social event.
Further, in step 1, news text data related to the social events are obtained from the news database, time, place, personnel and organization of the social events are obtained from the news text data, keywords of the social events are extracted through a TF-IDF algorithm, and synonyms and antonyms are added to the keywords.
Further, step 2 specifically includes the following steps:
step 201: and obtaining the time, the place, the personnel, the organization and the keyword information corresponding to each social event, and adding relations to different instances in the time, the place, the personnel, the organization and the keyword information according to the practical significance.
Step 202: the time, place, personnel, organization and key words of the social events are taken as objects, and the relationship of the object instances is used for constructing a heterogeneous information network expressed as
Wherein V represents an object of the object, wherein,
mapping a set of nodes to a set of objects for an object type mapping function, E representing a relationship, ψ E → R for a relationship type mapping function, mapping a set of edges to a set of relationships, meta-schema represented as T G The meta-schema T is a set of meta-schemas, where a represents a set of object types as nodes, object V is one object type of the set of object types a, R represents a set of relationship types, relationship E is one relationship type of the set of relationship types R, and meta-schema T is a set of meta-schemas G The object type set A is defined as a directed graph, the relation of the object type set A is an edge, and the number of the object types | A | > 1 or the number of the relation types | R | > 1 of the information network.
Further, step 3 specifically includes the following steps:
step 301: obtaining a plurality of meta-paths according to the time, the place, the personnel, the organization and the keyword information of the social events, wherein the meta-path P is expressed as
Wherein A is 1 、A 2 、A l+1 Each representing an object, R 1 、R 2 、R l Representing relationships between objects;
step 302: calculating the similarity between two social events through the meta-path:
wherein, Sim (e) i ,e j ) Representing the degree of similarity between two social events,
is a social event e i And e j Meta path P m Total number of examples above, w m Is a learnable parameter for measuring the importance of different meta-paths, b is a learnable bias parameter, ReLU (-) is an activation function, and the similarity between different social events is expressed as Sim (e) i ,e j )。
Further, step 4 specifically includes the following steps:
step 401: expressing each word in news text data related to the social event into a word vector through a pre-training language model BERT, inputting the word vector into a BiGRU model to calculate to obtain a hidden vector, and obtaining semantic features of the social event by using a fully-connected neural network;
Step 402: modeling the political influence, public opinion influence and emotional polarity of the social event as the social influence characteristic of the event;
step 403: splicing the semantic features and the social influence features to obtain fusion features of the social events;
further, step 4 specifically includes the following steps:
step 401 a: for a social event news text, each word is converted to a word vector X using the pre-trained language model BERT,
representing a matrix in which D w Is the dimension of the word vector, L s Is the length of the news text;
step 401 b: inputting the word vector of the ith word into a BiGRU model to calculate to obtain a hidden vector, wherein the hidden vector is expressed as:
wherein the hidden vector
Representing a vector
And
the splicing of (2) is carried out,
D h is the hidden layer dimension of the BiGRU model,
and
hidden states of i words, x, representing two directions i Is the ith line of the word vector X,
and
hidden states of i-1 words representing two directions;
step 401 c: obtaining semantic features f of social events using fully connected neural networks Sem Calculated according to the following formula:
where MLP represents a fully connected network and concat (. cndot.) represents a connect operation.
Step 402: for a social event, an evaluation value Score for political influence of the social event is obtained G News public opinion influence evaluation value Score I Event emotion polarity evaluation value Score T Wherein political influence estimates Score G Event emotion polarity evaluation value Score T Event emotion polarity assessment Score for scalar values obtainable from global news database GDELT I Calculated by the following formula:
Score I ＝concat([NumMentions；NumSources；NumArticles])
wherein NumMentations represents the news amount indirectly referring to the event in the global news database GDELT, NumSource represents the number of news media related to the event published in the global news database GDELT, and NumActicles represents the news media directly related to the event in the global news database GDELTNumber of news closed, social influence feature f Soc Calculated according to the following formula:
f Soc ＝MLP(concat[Score G ；Score I ；Score T ])
step 403: for a certain social event, the semantic features and the social influence features are spliced to obtain a fusion feature h of the social event, which is expressed as:
h＝concat([f Sem ；f Soc ])
the feature matrix H with dimension NxD is obtained by the feature representation of N social events, D is the vector length of the fusion feature H and is
Further, step 5 includes the following steps:
step 501: and taking the obtained similarity between the two social events as the edge weight between the nodes of the social event classification graph, and expressing the similarity as follows:
A[i,j]＝Sim(e i ,e j )
wherein A [ i, j ] represents the edge weight of the node i and the node j of the social event classification graph;
Step 502: and taking the social events as nodes, taking the obtained fusion characteristics H of the social events as characteristics of the nodes, taking the similarity between the two social events as edge weights between the nodes of the social event classification graph, and constructing the social event classification graph (A, H), wherein H is a characteristic matrix of news events serving as the nodes, and A is a matrix of the edge weights.
Further, step 6 includes the following steps:
step 601: constructing a social event classification model, wherein the social event classification model comprises a local aggregation graph neural network GraphSAGE layer, an automatic supervision graph attention network SuperGAT layer and a logic classification layer which are sequentially arranged;
inputting the social event classification graph into a local aggregation graph neural network GraphSAGE layer, wherein the characteristic vector output by a node i in the social event classification graph through the local aggregation graph neural network layer is represented as:
wherein h is i The characteristic h of the ith node is represented,
is the neighbor node set of the ith node in the graph, h j Features of neighbor nodes representing node i, W 1 ,W 2 A matrix of weight parameters that can be learned,
represents to be assembled
Characteristic vector h in j Adding, taking average value, and locally aggregating feature matrix H output by neural network layer GS Expressed as:
n is the total number of the nodes, 
Represents H GS The feature vector of the ith node, T represents a transposed matrix;
feature matrix H for outputting local aggregation map neural network layer GS Inputting into SuperGAT layer of attention network of self-supervision graph, and expressing as:
wherein the content of the first and second substances,
is the neighbor node set of the ith node in the graph, and the neighbor node set of the ith node
Any node j, α in j Weight coefficient, e, representing node j subjected to normalization j Represents the initial weight score of node j; LeakyReLU (-) is an activation function, exp (-) represents an exponential function. W 3 ,W 4 Is a learnable weight parameter matrix, d is a feature vector
Length of (d).
The feature matrix of the supervised graph attention network layer output is represented as:
n is the total number of the nodes,
represents H SG The feature vector of the ith node.
Feature matrix H for outputting attention network layer of self-supervision graph SG Input logical Classification layer, denoted as
Z＝sigmoid(H SG )
Wherein sigmoid is a classification function, and Z is a classification serial number of the output social event;
step 602: constructing a training set by the method in the step 1, respectively classifying social events of the training set according to event classes, setting class serial numbers for each event class, and performing training iteration on a social event classification model through data of the training set until the model converges to obtain a trained social event classification model;
Step 603: and inputting the social event classification graph corresponding to the social events needing to be classified into the social event classification model, outputting the classification serial numbers of the output social events needing to be classified, and obtaining the classification results of the social events through the classification serial numbers.
A computer device includes a memory, a processor, and a program stored on the memory and executable on the processor, wherein the processor executes the program to implement the social event classification method and apparatus based on the local aggregation graph attention network as described above.
A computer-readable storage medium on which a program is stored, characterized in that: the program is executed by a processor to realize the social event classification method and device based on the local aggregation graph attention network.
The invention has the beneficial effects that:
1. according to the method, the social events are modeled into the heterogeneous information network, the similarity between the two social events is calculated through the meta path, the method fully considers that five information dimensions of the social events are time, position, personnel, organization and keywords respectively, the complex social events are modeled into the heterogeneous information network with five dimensions, the event expression is more concise, the similarity measurement between the events is more accurate, and the interpretability is stronger.
2. The invention designs a set of meta-path set for measuring the similarity of social events, wherein the set comprises 25 meta-paths with different semantics, the internal association between the social events can be considered from multiple dimensions and aspects, and the idea and the reference are provided for the future similarity construction research based on the meta-paths.
3. The invention designs a new event feature extraction framework based on semantics and social influence, and combines the semantics feature with the social influence feature to obtain the mixed feature representation of the event. The semantic features firstly use a fully trained large-scale pre-training language model to obtain good word-level semantic representation, and then use a gated cycle unit (GRU) to fuse the integral actual meaning of a sentence, so that the semantic features used in the method integrate the specific semantics at the word level and the real meaning at the sentence level, and are beneficial to subsequent event classification; the social influence characteristic considers the political influence, the social public opinion influence and the emotional polarity of the event, models the social influence of the event from three dimensions, and lays a foundation for subsequent accurate classification.
3. The invention provides a novel local social event classification method, wherein a classification model uses a local aggregation graph attention network, the local aggregation graph attention network comprises a local aggregation graph neural network GraphSAGE layer and a self-supervision graph attention network SuperGAT layer, and both the local aggregation graph attention network and the self-supervision graph attention network SuperGAT layer have the advantages that the characteristics of neighbor nodes can be used without traversing all the nodes, the calculation complexity is greatly reduced, and the model training efficiency is improved. The local aggregation graph neural network uses the technologies of local sampling and neighbor aggregation, the local sampling enables the local aggregation graph neural network to cope with a large batch of graph data, and the neighbor aggregation technology enables the neighbor aggregation technology to reason the characteristics of newly added nodes; the self-supervision graph attention network can effectively deal with noise appearing in the network and can distribute more attention weight to neighbor nodes with consistent labels. Compared with the existing data classification models such as a Convolutional Neural Network (CNN), a gated recurrent neural network (GRU), a text graph convolutional network (TextGCN), a fast text classification method Fattext, a graph convolutional neural network (GCN), a graph attention network (GAT) and a large-scale pre-training language model BERT, the classification accuracy is remarkably improved.
4. The method can be used in the fields of social event analysis and data mining, can be particularly used for monitoring the prediction and supervision of a certain type of social events, and has wide application prospect.
Drawings
FIG. 1 is a schematic flow chart of the method and apparatus for social event classification based on a local aggregation graph attention network according to the present invention;
FIG. 2 is a schematic illustration of a meta-schema of a social event;
FIG. 3 is a diagram illustrating an internal structure of a computing device according to an embodiment.
Detailed Description
Referring to fig. 1, the social event classification method based on a local aggregation graph attention network of the present invention at least includes the following steps:
step 1: collecting news text data related to social events, and extracting time, place, personnel, organization and keyword information in the news text data;
step 2: modeling the time, place, personnel, organization and keyword information of the extracted social events into a heterogeneous information network, and extracting meta-modes of the heterogeneous information network;
and step 3: selecting a plurality of meta-paths with different semantics according to the meta-mode of the extracted heterogeneous information network, and calculating the similarity between two social events through the meta-paths;
And 4, step 4: obtaining semantic features of the social events from news text data related to the social events; modeling political influence, public opinion influence and emotional polarity of the social events into social influence characteristics of the social events, and splicing semantic characteristics of the social events and the social influence characteristics of the social events to obtain fusion characteristics of the social events;
and 5: using the similarity between the two obtained social events as the edge weight between the nodes of the social event classification graph, using the fusion characteristics of the obtained social events as the node characteristics, and constructing the social event classification graph;
and 6: and constructing and training a social event classification model, wherein the social event classification model comprises a local aggregation graph neural network GraphSAGE layer, a self-supervision graph attention network SuperGAT layer and a logic classification layer which are sequentially arranged, inputting the social event classification graph into the trained social event classification model, and outputting the category of the social event.
Specifically, in an embodiment of the present invention, in step 1, news text data related to a social event is obtained from a news database, in this embodiment, a global news database GDELT is selected to obtain the news text data, and event categories corresponding to the social event, such as collaboration, anti-meeting, fighting, and the like, are obtained, and in this embodiment, event category classifications as shown in table 1 are adopted:
| Serial number | Event category | Serial number | Event category |
| 1 | Disclosure statement | 11 | Disapproval of |
| 2 | (Appeal) | 12 | Rejection of |
| 3 | Intention to collaborate | 13 | Threat |
| 4 | Discussion of business | 14 | Anti-protocol |
| 5 | External cooperation | 15 | Dazzling military forces |
| 6 | Substantial cooperation | 16 | Reducing relationships |
| 7 | Assistance system | 17 | Stress |
| 8 | Compromise | 18 | Attack of |
| 9 | Investigation | 19 | Fight against |
| 10 | Require that | 20 | Large scale violence |
TABLE 1
Of course, in other embodiments of the present invention, more event category classifications may be used, and are not described herein.
After obtaining the number of news documents, the time, place, personnel and organization of the social events can be obtained from the news document data, and keywords of the social events are constructed, wherein the organization refers to a group or a group formed by mutually cooperating and combining people for realizing a certain target, such as party organization, workshop organization, enterprise organization, military organization and the like.
Specifically, in this embodiment, the keywords of the social event are extracted by the TF-IDF algorithm, and synonyms and antonyms are added to the keywords.
In this embodiment, step 2 specifically includes the following steps:
step 201: and acquiring the date, the position, the personnel, the organization and the keywords corresponding to each social event, and adding relations to different instances in the date, the position, the personnel, the organization and the keywords according to the practical significance.
Step 202: social events in the real world are typically extracted from news articles, which can be constructed as a graph with various types of nodes and edges, called heterogeneous information networks, in which certain types of objects connected by a sequence of relationships are often called meta-paths, which play an important role in capturing semantic correlations between objects in the heterogeneous information networks, and the present invention models social events based on the heterogeneous information networks and the meta-paths to obtain similarities between them.
The date, the position, the personnel, the organization and the key words of the social event are taken as objects, and the heterogeneous information network is constructed by the relationship of the object instance, wherein the information network is a function with object type mapping
E → R directed graph
Wherein V represents an object of the object, wherein,
mapping the set of nodes to a set of objects for an object type mapping function; e denotes a relationship, ψ E → R is a relationship type mapping function, the edge set is mapped to the relationship set, the meta-schema is denoted T G The relation E is a specific relation type in the relation type set R, and the element Mode T G The method comprises the following steps that (A, R) is a directed graph defined on an object type set A, the relation of the R is an edge, and the number of the object types | A | is larger than 1 or the number of the relation types | R | is larger than 1 of an information network.
By modeling the social events into a heterogeneous information network and calculating the similarity between the two social events through a meta path, the method fully considers that five information dimensions of the social events are time, position, personnel, organization and keywords respectively, and models the complex social events into the heterogeneous information network with five dimensions, so that the events are more concise to express, the similarity measurement between the events is more accurate and has stronger interpretability.
Further, step 3 specifically includes the following steps:
step 301: obtaining a plurality of meta-paths according to the date, the position, the personnel, the organization and the keyword information of the social events, wherein the meta-path P is expressed as
Wherein A is 1 、A 2 、A l+1 Each representing an object, R 1 、R 2 、R l Representing relationships between objects;
specifically, in this embodiment, 25 meta paths with different semantics are selected, see table 2.
TABLE 2
Wherein → denotes the path direction, in case of meta-path 5 "event 1 → person 2 → event 2", assuming that event 1 has three participants, student a, teacher B and school leader C, and event 2 has two participants, student a and teacher B, respectively, where a and B are in teacher-to-student relationship, a is the student of school leader C, and B is the teacher of school leader C, the total 4 instances on meta-path 5 with event 1 as the source node and event 2 as the target node are (1 → a → B → 2), (1 → B → a → 2), (1 → C → B → 2), and the total 6 instances on meta-path 5 with event 1 as the source node and target node: (1 → a → B → 1), (1 → a → C → 1), (1 → B → a → 1) (1 → B → C → 1), (1 → C → a → 1), (1 → C → B → 1), with 2 instances in total on the meta-path 5 of the event 2 source node and the target node: (2 → A → B → 2), (2 → B → A → 2).
In the embodiment, a set of meta-path sets for measuring the similarity of the social events is designed, wherein the set includes 25 meta-paths with different semantics, and the meta-path sets can consider internal association between the social events from multiple dimensions and aspects, so as to provide ideas and references for future similarity construction research based on the meta-paths.
Step 302: calculating the similarity between two social events through the meta-path:
wherein, Sim (e) i ,e j ) Representing the degree of similarity between two social events,
is a social event e i And e j Meta path P m Total number of examples above, w m Is a learnable parameter for measuring the importance of different meta-paths, b is a learnable bias parameter, ReLU (-) is an activation function, and the similarity between different social events is expressed as Sim (e) i ,e j )。
When j is equal to i, Sim is equal to 1. Although there is only one event, there are a variety of objects and relationships of events.
For example: taking the meta-path "event 1 → participant → event 2" as an example, event 1 is "student A thanks for teacher B", then there are two examples of the meta-path here: (1) event 1- > student a- > event 1(2) event 1- > teacher B- > event 1.
In this embodiment, step 4 specifically includes the following steps:
calculating semantic features of social events, comprising:
Step 401 a: for a social event news text, each word is converted to a word vector X using the pre-trained language model BERT,
representing a matrix in which D w Is the dimension of the word vector, L s Is the length of the news text;
step 401 b: inputting the word vector of the ith word into a BiGRU model to calculate to obtain a hidden vector, wherein the hidden vector is expressed as:
wherein the hidden vector
Representing a vector
And
the splicing of the two pieces of the paper is carried out,
D h is the hidden layer dimension of the BiGRU model,
and
hidden states of i words, x, representing two directions i Is the ith line of the word vector X,
and
hidden states of i-1 words representing two directions;
step 401 c: obtaining semantic features f of social events using fully connected neural networks Sem Calculated according to the following formula:
where MLP represents a fully connected network and concat (. cndot.) represents a connect operation.
Calculating social influence characteristics:
step 402: for a social event, a political influence evaluation value Score is obtained G News opinion influence evaluation value Score I Event emotion polarity evaluation value Score T Wherein political influence assessment value Score G Event emotion polarity evaluation value Score T Event emotion polarity assessment Score for scalar values obtainable from global news database GDELT I Calculated by the following formula:
Score I ＝concat([NumMentions；NumSources；NumArticles])
wherein NumNumNucations represents the news volume for indirectly mentioning the event in the global news database GDELT, NumSeurces represents the number of news media related to the event published in the global news database GDELT, NumAutics represents the news volume directly related to the event in the global news database GDELT, and the social influence characteristic f Soc Calculated according to the following formula:
f Soc ＝MLP(concat[Score G ；Score I ；Score T ])
where MLP represents a fully connected network and concat (. cndot.) represents a connect operation.
Calculating the fusion characteristics of the social events:
step 403: for a certain social event, the semantic features and the social influence features are spliced to obtain a fusion feature h of the social event, which is expressed as:
h＝concat([f Sem ；f Soc ])
the feature matrix H with dimension NxD is obtained by the feature representation of N social events, D is the vector length of the fusion feature H and is
Specifically, in step 5, the heterogeneous event classification problem is formulated as a graph-based node classification, and there are N social event news texts (S ═ S) in total 1 ,s 2 ,…,s N ). Each social event is a node. Edge weights A [ i, j ] between nodes]Calculated by step 3, the characteristics h of the node s Calculating by the step 4, wherein the social event classification graph is the basic input of the graph neural network, and the basic structure of the social event classification graph is the characteristics of nodes, edges and nodes; each node represents a social event, the node features consist of global context semantic features and local semantic features, and the edge weight between the two events is calculated by a similarity method based on knowledge.
In the embodiment, a new event feature extraction framework based on semantics and social influence is designed, and the semantic features and the social influence features are combined to obtain the mixed feature representation of the event. The semantic features firstly use a fully trained large-scale pre-training language model to obtain good word-level semantic representation, and then use a gated cycle unit (GRU) to fuse the integral actual meaning of a sentence, so that the semantic features used in the method integrate the specific semantics at the word level and the real meaning at the sentence level, and are beneficial to subsequent event classification; the social influence characteristic considers the political influence, the social public opinion influence and the emotional polarity of the event, models the social influence of the event from three dimensions, and lays a foundation for subsequent accurate classification.
The step 5 comprises the following steps:
step 501: and taking the obtained similarity between the two social events as the edge weight between the nodes of the social event classification graph, and expressing the similarity as follows:
A[i,j]＝Sim(e i ,e j )
wherein A [ i, j ] represents the edge weight of the node i and the node j of the social event classification graph;
step 502: and taking the social events as nodes, taking the obtained fusion characteristic H of the social events as the characteristic of the nodes, taking the similarity between the two social events as the edge weight between the nodes of the social event classification graph, and constructing the social event classification graph (A, H), wherein H is a characteristic matrix of the news events as the nodes, and A is a matrix of the edge weight.
The step 6 specifically comprises the following steps:
step 601: constructing a social event classification model, wherein the social event classification model comprises a local aggregation graph neural network GraphSAGE layer, an automatic supervision graph attention network SuperGAT layer and a logic classification layer which are sequentially arranged;
inputting the social event classification graph into a local aggregation graph neural network GraphSAGE layer, wherein the characteristic vector output by a node i in the social event classification graph through the local aggregation graph neural network layer is represented as:
wherein h is i The feature h of the ith node is represented,
is the neighbor node set of the ith node in the graph, h j Features of neighbor nodes representing node i, W 1 ,W 2 A matrix of weight parameters that can be learned,
represent a collection ofClosing box
Characteristic vector h in (1) j Adding the obtained values, and taking the average value to locally aggregate the feature matrix H output by the neural network GraphSAGE layer GS Expressed as:
n is the total number of the nodes,
represents H GS The feature vector of the ith node, T represents a transposed matrix;
feature matrix H for outputting local aggregation graph neural network GraphSAGE layer GS Inputting into SuperGAT layer of attention network of self-supervision graph, and expressing as:
wherein the content of the first and second substances,
is the neighbor node set of the ith node in the graph, and the neighbor node set of the ith node
Any node j, α in j Weight coefficient, e, representing node j subjected to normalization j Represents the initial weight score of node j; LeakyReLU (. circle.) is an activation function, exp (. circle.) tableAn index function is indicated. W 3 ,W 4 Is a learnable weight parameter matrix, d is a feature vector
Length of (d).
The feature matrix output by the SuperGAT layer of the supervision graph attention network is expressed as:
n is the total number of the nodes,
represents H SG The feature vector of the ith node.
Feature matrix H for outputting SuperGAT layer of self-supervision graph attention network SG Input logical Classification layer, denoted as
Z＝sigmoid(H SG )
Wherein sigmoid is a classification function, and Z is a category serial number of the output social event;
step 602: constructing a training set by the method in the step 1, respectively classifying social events of the training set according to event classes, setting class serial numbers for each event class, and performing training iteration on a social event classification model through data of the training set until the model converges to obtain a trained social event classification model;
step 603: and inputting the social event classification graph corresponding to the social events needing to be classified into the social event classification model, outputting the classification serial numbers of the output social events needing to be classified, and obtaining the classification results of the social events through the classification serial numbers.
Experimental demonstration is carried out on the social event classification method based on the local aggregation graph attention network in the embodiment:
the social event classification model based on the local aggregation graph attention network (GSSG: GraphSAGE + SuperGAT) and a Convolutional Neural Network (CNN), a gated recurrent neural network (GRU), a text graph convolutional network (TextGCN), a fast text classification method Fastext, a graph convolutional neural network (GCN), a graph attention network (GAT) and a large-scale pre-training language model BERT are compared with the event classification accuracy on a data set of five countries A, B, C, D and E, see table 3, the event classification accuracy data of the models on the data of the countries is given, and the experimental results show that the social event classification model has obvious superior performance in binary classification and twenty classification tasks.
TABLE 3
The invention provides a novel local social event classification method.A classification model uses a local aggregation graph attention network, which comprises a local aggregation graph neural network GraphSAGE layer and a self-supervision graph attention network SuperGAT layer, and both have the advantages that the characteristics of neighbor nodes can be used without traversing all the nodes, so that the calculation complexity is greatly reduced, and the model training efficiency is improved. The local aggregation graph neural network uses local sampling and neighbor aggregation technologies, the local sampling enables the local aggregation graph neural network to cope with a large batch of graph data, and the neighbor aggregation technology enables the neighbor aggregation technology to reason the characteristics of newly added nodes; the self-supervision graph attention network can effectively deal with noise occurring in the network and can distribute more attention weight to neighbor nodes with consistent labels. Compared with the existing data classification models such as a Convolutional Neural Network (CNN), a gated recurrent neural network (GRU), a text graph convolutional network (TextGCN), a fast text classification method (Fattext), a graph convolutional neural network (GCN), a graph attention network (GAT) and a large-scale pre-training language model BERT, the classification accuracy is remarkably improved.
In an embodiment of the present invention, a computer apparatus is further provided, which includes a memory and a processor, where the memory stores a computer program, and the processor executes the computer program to implement the social event classification method and apparatus based on the local aggregation graph attention network as described above.
The computer apparatus may be a terminal, and its internal structure diagram may be as shown in fig. 3. The computer device comprises a processor, a memory, a network interface, a display screen and an input device which are connected through a bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for connecting and communicating with an external terminal through a network. The computer program is executed by a processor to implement a method and apparatus for social event classification based on a local aggregation graph attention network. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on a shell of the computer device, an external keyboard, a touch pad or a mouse and the like.
The Memory may be, but is not limited to, a Random Access Memory (RAM), a Read Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Read-Only Memory (EPROM), an electrically Erasable Read-Only Memory (EEPROM), and the like. The memory is used for storing programs, and the processor executes the programs after receiving the execution instructions.
The processor may be an integrated circuit chip having signal processing capabilities. The Processor may be a general-purpose Processor, and includes a Central Processing Unit (CPU), a Network Processor (NP), and the like. The Processor may also be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
It will be appreciated by those skilled in the art that the configuration shown in fig. 3 is a block diagram of only a portion of the configuration associated with the present application, and is not intended to limit the computing device to which the present application may be applied, and that a particular computing device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.
In an embodiment of the present invention, a computer readable storage medium is further provided, on which a program is stored, and the program, when executed by a processor, implements the social event classification method and apparatus based on the local aggregation graph attention network as described above.
As will be appreciated by one of skill in the art, embodiments of the present invention may be provided as a method, computer apparatus, or computer program product. Accordingly, embodiments of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and so forth) having computer-usable program code embodied therein.
Claims (10)
1. A social event classification method based on a local aggregation graph attention network is characterized by comprising the following steps:
step 1: collecting news text data related to social events, and extracting time, place, personnel, organization and keyword information in the news text data;
step 2: modeling the time, place, personnel, organization and keyword information of the extracted social events into a heterogeneous information network, and extracting meta-modes of the heterogeneous information network;
and step 3: selecting a plurality of meta-paths with different semantics according to the meta-mode of the extracted heterogeneous information network, and calculating the similarity between two social events through the meta-paths;
and 4, step 4: obtaining semantic features of the social events from news text data related to the social events; modeling the political influence, the public opinion influence and the emotional polarity of the social events into social influence characteristics of the social events, and splicing the semantic characteristics of the social events and the social influence characteristics of the social events to obtain fusion characteristics of the social events;
and 5: using the similarity between the two obtained social events as the edge weight between the nodes of the social event classification graph, using the fusion characteristics of the obtained social events as the node characteristics, and constructing the social event classification graph;
And 6: and constructing and training a social event classification model, wherein the social event classification model comprises a local aggregation graph neural network GraphSAGE layer, a self-supervision graph attention network SuperGAT layer and a logic classification layer which are sequentially arranged, inputting the social event classification graph into the trained social event classification model, and outputting the category of the social event.
2. The method for classifying social events based on the local aggregation graph attention network as claimed in claim 1, wherein in step 1, news text data related to the social events is obtained from a news database, time, place, personnel and organization of the social events are obtained from the news text data, keywords of the social events are extracted through a TF-IDF algorithm, and synonyms and antonyms are added to the keywords.
3. The method of claim 1, wherein the social event classification method is based on a local aggregation graph attention network, and comprises the following steps: the step 2 specifically comprises the following steps:
step 201: obtaining time, place, personnel, organization and keyword information corresponding to each social event, and adding relations to different instances in the time, place, personnel, organization and keyword information according to the practical significance;
Step 202: the time, place, personnel, organization and key words of the social events are taken as objects, and the relationship of the object instances is used for constructing a heterogeneous information network expressed as
Wherein V represents an object of the group consisting of,
v → A is an object type mapping function, the node set is mapped to the object set, E represents a relationship, psi E → R is a relationship type mapping function, the edge set is mapped to the relationship set, and the meta-mode is represented as T G The meta-schema T is a set of meta-schemas, where a represents a set of object types as nodes, object V is one object type of the set of object types a, R represents a set of relationship types, relationship E is one relationship type of the set of relationship types R, and meta-schema T is a set of meta-schemas G The method comprises the following steps that (A, R) is a directed graph defined on an object type set A, the relation of the R is an edge, and the number of the object types | A | is larger than 1 or the number of the relation types | R | is larger than 1 of an information network.
4. The method for classifying social events based on the local aggregation graph attention network as claimed in claim 3, wherein the step 3 specifically comprises the following steps:
step 301: obtaining a plurality of meta-paths according to the time, the place, the personnel, the organization and the keyword information of the social events, wherein the meta-path P is expressed as
Wherein A is 1 、A 2 、A l+1 Each representing an object, R 1 、R 2 、R l Representing relationships between objects;
step 302: calculating the similarity between two social events through the meta-path:
wherein, Sim (e) i ,e j ) Representing the degree of similarity between two social events,
is a social event e i And e j Meta path P m Total number of examples above, w m Is a learnable parameter for measuring the importance of different meta-paths, b is a learnable bias parameter, ReLU (-) is an activation function, and the similarity between different social events is expressed as Sim (e) i ,e j )。
5. The method for classifying social events based on the local aggregation graph attention network as claimed in claim 4, wherein the step 4 specifically comprises the following steps:
step 401: expressing each word in news text data related to the social event into a word vector through a pre-training language model BERT, inputting the word vector into a BiGRU model to calculate to obtain a hidden vector, and obtaining semantic features of the social event by using a fully-connected neural network;
step 402: modeling political influence, public opinion influence and emotional polarity of the social events into social influence characteristics of the events;
step 403: and splicing the semantic features and the social influence features to obtain the fusion features of the social events.
6. The method for classifying social events based on the local aggregation graph attention network as claimed in claim 5, wherein the step 4 comprises the following steps:
Step 401 a: for a social event news text, each word is converted to a word vector X using the pre-trained language model BERT,
representing a matrix in which D w Is the dimension of the word vector, L s Is the length of the news text;
step 401 b: inputting the word vector of the ith word into a BiGRU model to calculate to obtain a hidden vector, wherein the hidden vector is expressed as:
wherein the hidden vector
Representing a vector
And
the splicing of the two pieces of the paper is carried out,
D h is the hidden layer dimension of the BiGRU model,
and
hidden states of i words, x, representing two directions i Is the ith line of the word vector X,
and
hidden states of i-1 words representing two directions;
step 401 c: obtaining semantic features f of social events using fully connected neural networks Sem Calculated according to the following formula:
wherein, MLP represents a fully connected network, concat (. cndot.) represents a connection operation;
step 402: for a social event, a political influence evaluation value Score is obtained G News opinion influence evaluation value Score I Event emotion polarity evaluation value Score T Wherein political influence assessment value Score G Event emotion polarity evaluation value Score T Event emotion polarity assessment Score for scalar values obtainable from global news database GDELT I Calculated by the following formula:
Score I ＝concat([NumMentions；NumSources；NumArticles])
Wherein NumNumNucations represents the news volume for indirectly mentioning the event in the global news database GDELT, NumSeurces represents the number of news media related to the event published in the global news database GDELT, NumAutics represents the news volume directly related to the event in the global news database GDELT, and the social influence characteristic f Soc Calculated according to the following formula:
f Soc ＝MLP(concat[Score G ；Score I ；Score T ])
step 403: for a certain social event, the semantic features and the social influence features are spliced to obtain a fusion feature h of the social event, which is expressed as:
h＝concat([f Sem ；f Soc ])
with dimension N x D derived from the feature representation of N social eventsThe feature matrix H, D is the vector length of the fusion feature H, is
7. The method for classifying social events based on the attention network of the local aggregation graph as claimed in claim 6, wherein step 5 comprises the following steps:
step 501: and taking the obtained similarity between the two social events as the edge weight between the nodes of the social event classification graph, and expressing the similarity as follows:
A[i,j]＝Sim(e i ,e j )
wherein A [ i, j ] represents the edge weight of the node i and the node j of the social event classification graph;
step 502: and taking the social events as nodes, taking the obtained fusion characteristic H of the social events as the characteristic of the nodes, taking the similarity between the two social events as the edge weight between the nodes of the social event classification graph, and constructing the social event classification graph (A, H), wherein H is a characteristic matrix of the news events as the nodes, and A is a matrix of the edge weight.
8. The method for classifying social events based on the local aggregation graph attention network as claimed in claim 6, wherein the step 6 comprises the following steps:
step 601: constructing a social event classification model, wherein the social event classification model comprises a local aggregation graph neural network GraphSAGE layer, an automatic supervision graph attention network SuperGAT layer and a logic classification layer which are sequentially arranged;
inputting the social event classification graph into a local aggregation graph neural network GraphSAGE layer, wherein the characteristic vector output by a node i in the social event classification graph through the local aggregation graph neural network layer is represented as:
wherein the content of the first and second substances,h i the feature h of the ith node is represented,
is the neighbor node set of the ith node in the graph, h j Features of neighbor nodes representing node i, W 1 ,W 2 A matrix of weight parameters that can be learned,
represents to be assembled
Characteristic vector h in j Adding the two layers, taking the average value, and locally aggregating a feature matrix H output by a neural network GraphSAGE layer of the graph GS Expressed as:
n is the total number of the nodes,
represents H GS The feature vector of the ith node, T represents a transposed matrix;
feature matrix H for outputting local aggregation graph neural network GraphSAGE layer GS Inputting into SuperGAT layer of attention network of self-supervision graph, and expressing as:
Wherein, the first and the second end of the pipe are connected with each other,
is the neighbor node set of the ith node in the graph, and the neighbor node set of the ith node
Any node j, α in (1) j Weight coefficient, e, representing node j subjected to normalization j Represents the initial weight score of node j; LeakyReLU (. circle.) is an activation function, exp (. circle.) represents an exponential function, W 3 ,W 4 Is a learnable weight parameter matrix, d is a feature vector
Length of (d);
the feature matrix output by the SuperGAT layer of the supervision graph attention network is expressed as:
n is the total number of the nodes,
represents H SG The feature vector of the ith node;
feature matrix H for outputting SuperGAT layer of self-supervision graph attention network SG Input logical Classification layer, denoted as
Z＝sigmoid(H SG )
Wherein sigmoid is a classification function, and Z is a classification serial number of the output social event;
step 602: constructing a training set by the method in the step 1, respectively classifying social events of the training set according to event classes, setting class serial numbers for each event class, and performing training iteration on a social event classification model through data of the training set until the model converges to obtain a trained social event classification model;
step 603: and inputting the social event classification graph corresponding to the social events needing to be classified into the social event classification model, outputting the classification serial numbers of the output social events needing to be classified, and obtaining the classification results of the social events through the classification serial numbers.
9. A computer device comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor when executing the program implements the method and apparatus for social event classification based on the local aggregation graph attention network according to claim 1.
10. A computer-readable storage medium on which a program is stored, characterized in that: the program when executed by a processor implements the method and apparatus for social event classification based on a local aggregated graph attention network as claimed in claim 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110787860.4A CN113449204B (en) | 2021-07-13 | 2021-07-13 | Social event classification method and device based on local aggregation graph attention network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110787860.4A CN113449204B (en) | 2021-07-13 | 2021-07-13 | Social event classification method and device based on local aggregation graph attention network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113449204A CN113449204A (en) | 2021-09-28 |
| CN113449204B true CN113449204B (en) | 2022-09-09 |
Family
ID=77816124
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110787860.4A Active CN113449204B (en) | 2021-07-13 | 2021-07-13 | Social event classification method and device based on local aggregation graph attention network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113449204B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114090902B (en) * | 2021-11-22 | 2022-09-09 | 中国人民解放军国防科技大学 | Social network influence prediction method and device based on heterogeneous network |
| CN115086004B (en) * | 2022-06-10 | 2023-08-29 | 中山大学 | Security event identification method and system based on heterogeneous graph |
| CN116303996B (en) * | 2023-05-25 | 2023-08-04 | 江西财经大学 | Theme event extraction method based on multifocal graph neural network |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110889434B (en) * | 2019-10-29 | 2023-04-21 | 东南大学 | Social network activity feature extraction method based on activity |
| CN111581983B (en) * | 2020-05-13 | 2023-08-08 | 中国人民解放军国防科技大学 | Method for predicting social concern hotspots in online public opinion event based on group analysis |
| CN112732905B (en) * | 2020-12-08 | 2022-06-14 | 银江技术股份有限公司 | Traffic accident analysis, prevention and control method and system based on knowledge graph |
| CN112749757B (en) * | 2021-01-21 | 2023-09-12 | 厦门大学 | Thesis classification model construction method and system based on gating graph annotation force network |
| CN112861967B (en) * | 2021-02-07 | 2023-04-07 | 中国电子科技集团公司电子科学研究院 | Social network abnormal user detection method and device based on heterogeneous graph neural network |
-
2021
- 2021-07-13 CN CN202110787860.4A patent/CN113449204B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN113449204A (en) | 2021-09-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Minh et al. | Deep learning approach for short-term stock trends prediction based on two-stream gated recurrent unit network | |
| Wang et al. | Deep learning for aspect-based sentiment analysis | |
| CN111159395B (en) | Chart neural network-based rumor standpoint detection method and device and electronic equipment | |
| CN113449204B (en) | Social event classification method and device based on local aggregation graph attention network | |
| CN111143576A (en) | Event-oriented dynamic knowledge graph construction method and device | |
| CN109726745B (en) | Target-based emotion classification method integrating description knowledge | |
| Fu et al. | Listening to the investors: A novel framework for online lending default prediction using deep learning neural networks | |
| Katz et al. | Natural language processing in the legal domain | |
| CN105354216A (en) | Chinese microblog topic information processing method | |
| CN111368082A (en) | Emotion analysis method for domain adaptive word embedding based on hierarchical network | |
| Golestani et al. | Real-time prediction of employee engagement using social media and text mining | |
| CN113312480A (en) | Scientific and technological thesis level multi-label classification method and device based on graph convolution network | |
| CN113742733A (en) | Reading understanding vulnerability event trigger word extraction and vulnerability type identification method and device | |
| Fu et al. | Improving distributed word representation and topic model by word-topic mixture model | |
| Shrivastava et al. | Enhancing aggression detection using GPT-2 based data balancing technique | |
| Shi et al. | EKGTF: A knowledge-enhanced model for optimizing social network-based meteorological briefings | |
| Domingos et al. | Just add weights: Markov logic for the semantic web | |
| Wang et al. | Automatic dialogue system of marriage law based on the parallel C4. 5 decision tree | |
| Bi et al. | Judicial knowledge-enhanced magnitude-aware reasoning for numerical legal judgment prediction | |
| CN111859955A (en) | Public opinion data analysis model based on deep learning | |
| Ray et al. | Prediction and Analysis of Sentiments of Reddit Users towards the Climate Change Crisis | |
| Zhu et al. | Attention based BiLSTM-MCNN for sentiment analysis | |
| Jiang et al. | Sentiment classification based on clause polarity and fusion via convolutional neural network | |
| Dutta et al. | Fake news prediction: a survey | |
| Bestvater | Using machine learning to infer real-world political attitudes and behaviors from social media data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |