Event knowledge graph GCN embedding representation method, system, device and terminal
By constructing a GCN event representation model with behavioral situation semantics, the problem of low accuracy of event classification in the prior art is solved, and high accuracy recognition and deep semantic representation of specific scene events are achieved, which enhances the credibility of event knowledge.
Patent Information
- Application Number
- CN202310631121.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-31
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-05-31
AI Technical Summary
The existing event knowledge graph extraction technology in specific scenarios caused by the diversity of natural language to be unable to distinguish event behaviors of the same meaning due to the diversity of natural language, resulting in low accuracy of event classification and prediction tasks, and lack of form for event and related entity states. .
By constructing the behavioral base GCN event representation model of the state layer-behavior layer-event layer of behavioral situation semantics, using the embedding of state node entities and node relationships, an interpretable expression of the behavioral state space of action-driven events to event association is formed, reducing the cost of Reliance on lexical grammar rules.
It improves the accuracy of behavior classification recognition in specific scenario events, enhances the credibility of event knowledge, provides deep semantic information of events and a basis for verifying event authenticity, laying the foundation for event prediction and reasoning tasks.
Smart Images

Figure CN116629358B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of event knowledge representation, and in particular, relates to an event knowledge graph GCN embedding representation method, system, device and terminal. Background Art
[0002] The event knowledge graph is an event-centric knowledge graph used to describe event information and various relationships between events. Its main research object is events, including knowledge related to events, the evolution of events, and the correlation between events. In the process of constructing, reasoning, and applying event graphs, a variety of information processing technologies and knowledge representation methods are needed. The core technology at the upstream is event extraction technology. Event extraction technology extracts event-related information from unstructured text data and presents the information in a structured form, laying the foundation for downstream tasks such as event prediction and event reasoning. Existing research mainly has the following problems:
[0003] 1. Loss of semantic information
[0004] In the field of event knowledge representation, most existing studies, represented by ACE, believe that the trigger word of an event is the word that most clearly expresses its occurrence, and regard verbs as the most important components, but lack a formal representation of the state description of events and related entities, which may lead to information loss. Due to the ambiguity of natural language, the mapping accuracy from the identified trigger words to the corresponding action-driven event types is limited. Modeling events based on verb phrases only at the lexical and grammatical concept level cannot verify the authenticity of the event.
[0005] 2. Insufficient ability to recognize situational semantics
[0006] At present, the research on event knowledge graph extraction technology mainly adopts deep learning methods. Representative neural network methods include core word extraction methods based on attention mechanism, word vector entity representation methods based on pre-trained models, and entity and relationship fusion extraction methods based on graph neural networks. These three methods rely on relevant verb lexical or grammatical rules to identify causal or temporal associations between events. However, in specific scenarios, due to the diversity of natural languages, behaviors or actions are not expressed using the same or similar verbs, resulting in verb lexical or grammatical rules being unable to distinguish between events with the same meaning, making the accuracy of event classification and prediction tasks based on trigger words lower than expected.
[0007] Therefore, it is necessary to construct a formal representation model and related embedding technology that has weak dependence on lexical and grammatical rules, has good recognition for specific scenarios, and is more expressive for complex events composed of multiple events. The present invention aims to solve the above problems and proposes a GCN embedding representation method for event knowledge graph based on behavioral situation semantics. Summary of the invention
[0008] In response to the problems existing in the prior art, the present invention provides a method, system, device and terminal for embedding an event knowledge graph GCN, and in particular, relates to a method, system, medium, device and terminal for embedding an event knowledge graph GCN based on behavioral situation semantics.
[0009] The present invention is implemented as follows: a GCN embedding representation method for an event knowledge graph, the event knowledge graph GCN embedding representation method includes: establishing a state information space for event behaviors in a specific scene, interpreting and defining the situation semantics of behavior-related state information changes through a state set; then representing the state layer of the behavior situation semantics as GCN nodes and edges, constructing a behavior base through the embedding of state node entities and node relationships, and forming a state layer-behavior layer-event layer behavior base GCN event representation model, specifically including the following steps:
[0010] Step 1: Data annotation: predefine events for each conversation data and annotate the behavior state space attributes, and convert them into a behavior state space structured data set;
[0011] Step 2: word vector model training. According to step 1, the behavior types are defined for the training data set and the test data set, and the word vector model is trained on the annotated content of the state space entity to generate a word vector representing the annotated data set.
[0012] Step 3: Build a behavioral GCN based on the vector file of the word vector pair. According to step 2, input the vector file content to the GCN construction algorithm and output the behavioral basis of each dialogue data annotation dataset.
[0013] Furthermore, the data annotation in step 1 includes:
[0014] (1) Define events and behavior state space attributes for each conversation;
[0015] (2) Refine the discrete function value domain space of behavioral situation semantics and expand the state information entity.
[0016] Furthermore, in step 1, six attributes of event representation are identified for each conversation, including conversation ID, content, behavior, object, state, and category.
[0017] Furthermore, the word vector model training in step 2 includes:
[0018] (1) Using the Word2Vector algorithm, we define the behavior type for the test dataset and train the word vector model for the annotation content of the state space entity.
[0019] (2) Using Skip-grams network training parameters to transform entities in the data state space into word vectors;
[0020] (3) After processing, a word vector representing the optimized labeled data set is generated.
[0021] Furthermore, the construction of GCN in step 3 includes:
[0022] (1) Construct a vector file of word vector pairs based on the behavior type state space graph, and store the node and node pair information in the GCN graph;
[0023] (2) Input the GCN construction algorithm and perform GCN embedding;
[0024] (3) Output the behavioral basis of the labeled dataset dialogue.
[0025] Another object of the present invention is to provide an event knowledge graph GCN embedding representation system using the event knowledge graph GCN embedding representation method, the event knowledge graph GCN embedding representation system comprising:
[0026] The data annotation module is used to predefine events and annotate the behavior state space attributes for each conversation data, and convert it into a behavior state space structured data set;
[0027] The word vector model training module is used to define the behavior type for the training data set and the test data set, train the word vector model for the annotated content of the state space entity, and generate the word vector representing the annotated data set;
[0028] The GCN construction module is used to construct a vector file of word vector pairs based on the behavior type state space graph, input the nodes and node pair information in the GCN graph to the GCN construction algorithm, and output the behavior basis of each dialogue data annotation dataset.
[0029] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the event knowledge graph GCN embedding representation method.
[0030] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the event knowledge graph GCN embedding representation method.
[0031] Another object of the present invention is to provide an information data processing terminal, which is used to implement the event knowledge graph GCN embedding representation system.
[0032] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:
[0033] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving the problems, the technical solutions to be protected by the present invention and the results and data during the research and development process are closely combined to analyze in detail and deeply how the technical solutions of the present invention solve the technical problems, and some creative technical effects brought about after solving the problems. The specific description is as follows:
[0034] The present invention provides a GCN embedding representation method for event knowledge graph based on behavior situation semantics. By constructing an event behavior state space and situation semantic representation model in a specific scenario, the state layer of the behavior situation semantics is represented as nodes and edges of GCN, and then the behavior base is constructed by embedding state node entities and node relationships, thereby forming an interpretable expression of the behavior state space of action-driven events to event associations.
[0035] In addition, the present invention also has the following beneficial effects:
[0036] 1. In many previous event knowledge graph analyses, the identification of causal or temporal associations between events depends on the relevant verb lexical or grammatical rules. However, in specific scene events, behaviors or actions are not expressed using the same or similar verbs, resulting in event classification and prediction tasks based on verb lexical or grammatical rules being less accurate than expected. Therefore, the technical solution of the present invention forms an interpretable expression of the behavior state space of action-driven events to event associations by constructing an event behavior state space and situation semantic representation model in a specific scene and applying the GCN embedding method, thereby improving the accuracy of behavior classification and recognition in specific scene events.
[0037] 2. Currently, there is no model for event recognition and classification based on behavioral situation semantics and GCN embedding in the prior art. The present invention establishes a state information space for event behaviors in specific scenarios, interprets and defines situation semantics for behavior-related state information changes through state sets, and proposes a behavior-based GCN event representation model of state layer-behavior layer-event layer. Based on GCN's ability to extract feature attributes of graph structures and the characteristics of layer-by-layer attribute transmission, the state layer of behavioral situation semantics is represented as GCN nodes and edges, and a behavior base is constructed by embedding state node entities and node relationships, thereby providing behavioral attribute calculation for the behavior layer and event layer, and realizing the recognition and classification of events in specific scenarios.
[0038] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by the present invention are described in detail as follows:
[0039] Based on the behavior situation semantics theory, the present invention proposes a behavior knowledge representation method and embedding method of event knowledge graph, establishes a state information space for event behavior in a specific scene, interprets and defines the behavior-related state information changes through state sets, and constructs a behavior-based GCN event representation model of state layer-behavior layer-event layer, thereby forming an interpretable expression of the behavior state space of action-driven events to event associations. The method proposed by the present invention provides information conversion of entities involved in the event, identifies the occurrence of the event through information conversion, can be used as a basis for verifying the authenticity of the event, and also enriches the deep semantic information of the event, laying the foundation for downstream tasks such as event prediction and event reasoning.
[0040] The behavior-based GCN proposed in the present invention integrates graph convolutional networks and behavior situation semantics formalization methods, uses behavior functions to define situation semantics, and constructs graph-based semi-supervised learning. The present invention constructs an event behavior state space in a specific scenario, represents the state layer of behavior situation semantics as nodes and edges of GCN, and then constructs a behavior base by embedding state node entities and node relationships, providing a knowledge representation embedding method at the situation semantics level for event extraction tasks, improving the accuracy of behavior classification and recognition in specific scenario events, and enhancing the credibility of event knowledge in context.
[0041] Third, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:
[0042] (1) The technical solution of the present invention fills the technical gap in the industry at home and abroad:
[0043] The current event knowledge graph has different advantages and disadvantages in its knowledge representation methods for event-related entity information: ACE 2005 regards the trigger of an event as the vocabulary that best expresses the event. Its advantage is that it gives a strong definition of the event, but its disadvantage is that the mapping from the recognized verb to the corresponding event type may lead to information loss; NEEG uses a neural network structure to propose a more expressive knowledge representation model. Its advantage is that neural networks are suitable for encoding, but its disadvantage is that it only models events in a conceptual sense and still regards verbs as the most important components; ESO formalizes the postconditions of events and the roles of entities related to the event, but lacks a formal representation of the state description of events and related entities.
[0044] The behavioral knowledge representation method and embedding method of the event knowledge graph based on behavioral situation semantics proposed in the present invention establishes a state information space for event behaviors in specific scenarios, uses state sets to perform situation semantic interpretation and definition on behavior-related state information changes, and constructs a behavior-based GCN event representation model of state layer-behavior layer-event layer. This provides a knowledge representation embedding method at the situation semantic level for event extraction tasks, enriches the deep semantic information of events, and lays the foundation for downstream tasks such as event prediction and event reasoning.
[0045] (2) The technical solution of the present invention solves a technical problem that people have been eager to solve but have never been able to solve successfully:
[0046] At present, the research on event extraction technology mainly adopts deep learning methods. Representative neural network methods include core word extraction methods based on attention mechanisms, word vector entity representation methods based on pre-trained models, and entity and relationship fusion extraction methods based on graph neural networks. These three methods rely on relevant verb lexical or grammatical rules to identify causal or temporal associations between events. However, in certain scenarios, behaviors or actions are not expressed using the same or similar verbs, resulting in verb lexical or grammatical rules being unable to distinguish between events with the same meaning, making the accuracy of event classification and prediction tasks based on trigger words lower than expected.
[0047] One of the urgent needs in the current field of event knowledge representation is to build a formal expression model and related embedding technology that has weak dependence on lexical and grammatical rules, has good recognition for specific scenarios, and is more expressive for complex events composed of multiple events.
[0048] The behavior-based GCN proposed in the present invention integrates graph convolutional networks and behavioral situation semantic formalization methods, constructs a behavior state space for events in specific scenarios, interprets and defines situation semantics for behavior-related state information changes through state sets, provides information conversion of entities involved in events, and identifies the occurrence of events through information conversion, which can serve as a basis for verifying the authenticity of events, enhances the credibility of event knowledge in context, and improves the accuracy of behavior classification and recognition in specific scenario events. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0050] Figure 1It is a flow chart of the event knowledge graph GCN embedding representation method provided by an embodiment of the present invention;
[0051] Figure 2 It is a schematic diagram of the event knowledge graph GCN embedding representation method provided by an embodiment of the present invention;
[0052] Figure 3 It is a three-layer event formalization model diagram provided by an embodiment of the present invention;
[0053] Figure 4 It is a three-layer event representation model diagram provided by an embodiment of the present invention;
[0054] Figure 5 It is a six-type behavior-based GCN graph provided by an embodiment of the present invention;
[0055] Figure 6 It is a Word2Vector model diagram provided by an embodiment of the present invention;
[0056] Figure 7 This is a flow chart of the behavior GCN construction provided by an embodiment of the present invention.
[0057] Figure 8 is an example of labeled data provided by an embodiment of the present invention;
[0058] Fig. 9 It is a vector file instance and content example provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0060] In view of the problems existing in the prior art, the present invention provides an event knowledge graph GCN embedding representation method, system, device and terminal. The present invention is described in detail below with reference to the accompanying drawings.
[0061] like Figure 1 As shown, the event knowledge graph GCN embedding representation method provided by the embodiment of the present invention includes the following steps:
[0062] S101, constructing the event behavior state space and situation semantic representation model in a specific scenario;
[0063] S102, representing the state layer of the behavior situation semantics as nodes and edges of GCN;
[0064] S103, constructing a behavior base by embedding state node entities and node relationships, forming an interpretable expression from the behavior state space to the event association.
[0065] As a preferred embodiment, Figure 2 As shown, the event knowledge graph GCN embedding representation method provided by the embodiment of the present invention specifically includes the following steps:
[0066] S1. Data annotation: predefine events for each conversation data and annotate the behavior state space attributes, and convert it into a behavior state space structured data set;
[0067] S2, word vector model training, according to step S1, define the behavior type for the training data set and the test data set, perform word vector model training on the annotated content of the state space entity, and generate a word vector representing the annotated data set;
[0068] S3. Build a behavioral GCN based on the vector file of the word vector pair. According to step S2, input the vector file content to the GCN construction algorithm and output the behavioral basis of each dialogue data annotation dataset.
[0069] The principle of the event knowledge graph GCN embedding representation method provided by the embodiment of the present invention is:
[0070] By constructing an event behavior state space and situation semantic representation model in a specific scenario, the event is formally defined using the behavior basis, and the GCN graph convolution structure is used to embed the events and the relationships between them into a continuous vector space. Finally, the situation semantics of the event is extracted, thus forming an interpretable expression of the behavior state space of action-driven events to event associations, which improves the accuracy of behavior classification and recognition in specific scenario events.
[0071] The present invention is based on behavioral situation theory and situation semantics. The behavioral situation is the meaning of the behavior operation. By using the behavioral situation semantics formalization method, the state information set of observable behaviors is taken as the object of situation, and the situation meaning and expression method of its information are analyzed. Given the relevant environment state space:
[0072] E S =Env ST ×Sub ST
[0073] Behavior situation semantics is a function based on this state set, which can be expressed as a set E S To set E S The mapping is:
[0074] F situation :E S →E S
[0075] The present invention interprets the behavior situation semantics as a function of the state information set by extracting relevant state information during the event process. st ×Obj st It is a state set that establishes a state space for the subject and object in the event. The behavior extraction function is the interpretation function of the situation semantics when the behavior occurs. It is a mapping from I to I, that is, B:I→I, which represents the state transition of the subject and object caused by the running behavior.
[0076] The embodiment of the present invention proposes a behavior-based GCN event representation model based on the state layer-behavior layer-event layer according to the behavior situation semantic model and the related behavior set in a specific scenario, such as Figures 3-4 As shown, the detailed mechanism of each layer is as follows:
[0077] State layer: The state layer refers to the information transformation function between entities involved in an event, which can be considered as observable information of the behavior:
[0078] B:I→I,I=Info sub ×Info obj
[0079] Action layer: The action layer can be regarded as a set of actions of the event subject, each action represents an information conversion function:
[0080] Act={A1,A2,..,A n},A i =b|b·b,b∈B
[0081] Event layer: The event layer can be composed of atomic events or compound events. An atomic event is the core event or indivisible event of a complete event, and a compound event is an event with behavioral associations. The formal representation of an atomic event is a five-tuple:
[0082] E={Sub,Act,Obj,Time,Loc}
[0083] The embodiment of the present invention represents the state layer of behavior situation semantics as nodes and edges of GCN, and uses GCN's ability to extract feature attributes of graph structures and the characteristics of layer-by-layer attribute transmission to build a behavior base by embedding state node entities and node relationships, thereby providing behavior attribute calculation for the behavior layer and event layer, and realizing the recognition and classification of events in specific scenarios. The GCN model is a general term for graph neural networks that use filters with shared parameters, and its input is the node feature matrix X N×D and the adjacency matrix A describing the graph structure N×N , after transformation, the output feature matrix Z N×F , where N is the number of nodes, D is the input feature dimension, and F is the output feature dimension, or the embedded dimension.
[0084] Each layer of propagation in GCN is a nonlinear equation:
[0085] H (l+1) =f(H (l) ,A)
[0086] Among them, l is the number of layers of GCN, H (0) =X,H (l) = Z. The difference between different GCN models lies in the choice of nonlinear equation f(·,·) and how to set the parameters of this equation. Usually, you can choose to use a nonlinear activation function, such as ReLU.
[0087] The behavior-based GCN in the embodiment of the present invention refers to the content conversion function of the state entity related to the contextual content information involved in the event. In the behavior-driven event, the state layer constructs the behavior, where B:I→I represents the behavior situation semantics. The embodiment of the present invention uses the potential state space of event information, represented as I=Info sub ×Info obj , starting from the change of the state of the event subject and object caused by the occurrence of the event, combined with the GCN model, the event state layer is embedded in GCN, and the concept of behavior base is proposed:
[0088] Let I = Info sub ×Info obj As a set of behavioral states, i∈I is a node of GCN, and the state transition relationship between i is the edge of GCN. The constructed GCN is a graph mapping from I to I, that is, B:I→I. This GCN is called a behavioral basis.
[0089] The behavior base is the basis of the behavior-driven event, that is, the state transition function of the entity involved in the event. The specific definition of the behavior base varies depending on the event type. For transaction events in shopping scenarios, the embodiment of the present invention defines six types of behavior-driven events:
[0090] E={e order ,e return ,e exchange ,e receive ,e payment ,e deliver}
[0091] The driving behavior bases of the above six types of events are defined as:
[0092] B={B order ,B return ,B exchange ,B receive ,B payment ,B deliver}
[0093] The GCN diagram of these six types of behavior bases is as follows Figure 5 shown.
[0094] The embodiment of the present invention uses the Word2Vec word vector algorithm to construct the GCN graph constructed by the relationship between the behavior-based related nodes and the edges, and constructs the behavior-based nodes representing the state space into word vectors. The Word2Vec word vector model is an unsupervised text corpus training model, and its basic structure diagram is as follows Figure 6 It is a distributed representation of words, using a fixed-length vector to represent words. This vector also contains the semantic information of the words. In the input layer of the Word2Vec model, the input received is the one-hot vector x of the word. After the shared weight matrix W V×N Get the hidden layer vector h, that is:
[0095] h=x·W V×N
[0096] Then through the weight matrix W' V×N Get the prediction of the output word, that is:
[0097] y=h·W' V×N
[0098] In the embodiment of the present invention, the focus is not on the prediction of the word context, but on the byproduct generated during the training process of this task, namely, the model parameters. In the model parameters, each vector corresponds to a word, and these parameters are used as the word vector representation of the word.
[0099] The behavior-based GCN model in the present invention is:
[0100] The present invention defines the GCN embedding method of behavior situation semantics as the construction process of the state graph and the construction process of the adjacency matrix, such as Figure 7 As shown. The behavioral state diagram can be defined as:
[0101] G={V,E,C}
[0102] Among them, V is the point set; E is the edge set; C is the category set, which explains the category of the current graph. V = {S, I, W, L}, S is the state set, I is the node index, L is the category set, which is used to describe the category of the node, and W is the node word vector, which represents the node vector corresponding to the state entity, obtained by the Word2Vec model trained after the original corpus is annotated with state information.
[0103] In the embodiment of the present invention, the original data set is divided into data in the form of dialogue id. The dialogue id is used as the processing unit of data input, which describes the relevant events and the relevant behaviors that drive the events. Therefore, the embedding algorithm uses the dialogue id as the original input corpus to construct the state diagram and transition matrix of the corresponding behavior base for each group of dialogues. Thus, k∈M is defined, k={1,…,M}, M is the number of dialogue ids, and i∈N, i={1,…,N k},N k is the number of nodes in the behavioral base state graph.
[0104] In the behavior base state diagram, suppose i∈N k Represents a node, a node pair express and The edges between them, the adjacency matrix A k Indicated as N k ×N k ,in, Used to determine the existence of edges. Label the relevant nodes and node pairs in the form of node names in each input conversation id to provide valid input for the embedding algorithm. The embedding algorithm calls the behavior base state graph construction function to generate a behavior base for each conversation id.
[0105] The adjacency matrix A of the behavior basis k For collection Where k∈M, i,j∈N.
[0106] In the node set V, i corresponds to all node numbers, C corresponds to the category of the behavior state diagram, and W is the word vector of all nodes. L is the category of the node. For each node V i The relevant information is stored in the node file, providing subsequent data query for adjacency matrix conversion and subsequent calculations.
[0107] The implementation method of the technical solution of the present invention is:
[0108] 1. Data annotation stage
[0109] In order to obtain a structured data set of behavior state space, events are predefined for each dialogue data and behavior state space attributes are annotated. That is, six attributes are identified for each dialogue: dialogue ID, content, behavior, object, state, and category. Based on the preliminary annotation, the discrete function value domain space of the behavior situation semantics is further refined to expand the "state" information entity.
[0110] 2. Word vector model training stage
[0111] First, the Word2vector algorithm is used to define the behavior type of the test dataset, and the word vector model is trained for the annotation content of the state space entity. Then, the skip-grams network training parameters are used to convert the project entities in the data state space into word vectors. After processing, the word vector representing the optimized annotated dataset is generated.
[0112] 3. GCN construction stage
[0113] Construct a vector file of word vector pairs based on the behavior type state space graph, store the nodes and node peer information in the GCN graph, input the GCN construction algorithm, perform GCN embedding, and output the behavior basis of the labeled dataset dialogue.
[0114] 4. Test results
[0115] The tool used in the present invention has completed a series of complete construction processes such as relevant data annotation, model training and system testing in the field of event knowledge representation technology, and has been run under real data.
[0116] Through the embodiments of the present invention, the technical solution of the present invention establishes a state information space for event behaviors in specific scenarios, and explains the occurrence of behaviors under specific events; and based on the ability of GCN to extract feature attributes of graph structures and the characteristics of layer-by-layer attribute transmission, a behavior base is constructed by embedding state node entities and node relationships, which can form an interpretable expression of the behavior state space of action-driven events to event associations, thereby providing behavior attribute calculation for the behavior layer and the event layer, and improving the accuracy of behavior classification and recognition in specific scenario events.
[0117] The event knowledge graph GCN embedding representation system provided by the embodiment of the present invention includes:
[0118] The data annotation module is used to predefine events and annotate the behavior state space attributes for each conversation data, and convert it into a behavior state space structured data set;
[0119] The word vector model training module is used to define the behavior type for the training data set and the test data set, train the word vector model for the annotated content of the state space entity, and generate the word vector representing the annotated data set;
[0120] The GCN construction module is used to construct a vector file of word vector pairs based on the behavior type state space graph, input the nodes and node pair information in the GCN graph to the GCN construction algorithm, and output the behavior basis of each dialogue data annotation dataset.
[0121] In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.
[0122] For the GCN embedding method of the event knowledge graph behavior knowledge representation based on behavior situation semantics proposed in the present invention, this embodiment mainly considers the following factors when selecting corpus data:
[0123] (1) It should be corpus in a specific scenario, corresponding to the setting of “situational semantics” in this invention;
[0124] (2) The scenario should be somewhat common;
[0125] (3) The corpus should be of a certain size to ensure the data reliability for group training and verification.
[0126] Considering the above factors, this embodiment selects the chat records between JD merchants and customers as the original corpus and training data set.
[0127] The event knowledge graph GCN embedding representation method provided in this embodiment specifically includes the following steps:
[0128] Step 1: Convert the original data set into a structured data set in the behavior state space. First, label the relevant data, predefine events for each conversation data, and label the behavior state space attributes, that is, identify six basic attributes for each conversation, including conversation ID, content, behavior, object, state, and category. Further, refine the discrete function value domain space of the behavior situation semantics, that is, expand the "state" information entity. The labeled data instance is as follows: Figure 8 shown.
[0129] Step 2: Word vector model training. Based on step 1, the word2vector algorithm is used to define the behavior types for the training data set and the test data set, and the word vector model is trained for the labeled content of the state space entity. The skip-grams network training parameters are used to transform the project entities in the data state space into word vectors to generate the following: Fig. 9 The five vector files shown represent the word vectors generated to represent the optimized labeled dataset.
[0130] Step 3: Build a vector file of word vector pairs based on the behavior type state space graph, store the nodes and node peer information in the GCN graph, input the GCN construction algorithm, perform GCN embedding, and output the behavior basis of the labeled dataset dialogue.
[0131] At present, the advanced model methods in the field of event extraction mainly include the following categories:
[0132] (1) The Cross Event method uses the entire document level information for event extraction;
[0133] (2) DMCNN established a dynamic multi-pooling CNN model;
[0134] (3) JRNN uses bidirectional RNN and artificial input features;
[0135] (4) DEEB-RNN uses hierarchical, supervised attention mechanism and document-level information for event extraction;
[0136] (5) dbRNN adds grammatical dependency arcs to the bidirectional LSTM network to enhance event extraction capabilities;
[0137] (6) GCN-ED uses a parameter pool mechanism and performs event extraction based on GCN;
[0138] (7) JMEE uses syntactic shortcut arcs and attention-based GCN to extract event information.
[0139] The event extraction accuracy of the above methods on the ACE 2005 test dataset is shown in Table 1.
[0140] Table 1 Event extraction accuracy of different methods on ACE 2005 test dataset
[0141] method Accuracy (%) CrossEvent 68.7 DMCNN 75.6 JRNN 66.0 DEEB-RNN 72.3 dbRNN 74.1 GCN-ED 77.9 JME 76.3
[0142] The above methods cannot accurately extract events based on scenarios and context models to meet the needs of downstream tasks. For example, when processing JD chat data, the accuracy rate does not exceed 60%.
[0143] This embodiment is based on situational semantics, so the chat records between JD merchants and customers are selected as the original corpus and training data set. The actual recognition accuracy of the sample set is shown in Table 2. In specific scenarios, the accuracy of behavior classification recognition has reached an advanced level for atomic events and complex events composed of multiple events.
[0144] Table 2 Event extraction accuracy of the embodiment of the present invention in the test set
[0145]
[0146]
[0147] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.
[0148] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.
Claims
1. A GCN embedding representation method for event knowledge graph, characterized in that: The event knowledge graph GCN embedding representation method includes: establishing a state information space for event behaviors in a specific scenario, interpreting and defining the situational semantics of behavior-related state information changes through state sets; then representing the state layer of behavior situation semantics as GCN nodes and edges, and constructing a behavior base through embedding state node entities and node relationships, forming a behavior base GCN event representation model of state layer-behavior layer-event layer; The event knowledge graph GCN embedding representation method includes the following steps: Step 1: Data annotation: predefine events for each conversation data and annotate the behavior state space attributes, and convert them into a behavior state space structured data set; Step 2: word vector model training. According to step 1, the behavior types are defined for the training data set and the test data set, and the word vector model is trained on the annotated content of the state space entity to generate a word vector representing the annotated data set. Step 3: Build a behavioral GCN based on the vector file of the word vector pair. According to step 2, input the vector file content to the GCN construction algorithm and output the behavioral basis of each dialogue data annotation dataset. The word vector model training in step 2 includes: (1) Using the Word2Vector algorithm, we define the behavior type for the test dataset and train the word vector model for the annotation content of the state space entity. (2) Using Skip-grams network training parameters to transform entities in the data state space into word vectors; (3) After processing, a word vector representing the optimized labeled data set is generated; Data annotation in step 1 includes: (1) Define events and behavior state space attributes for each conversation; (2) Refine the discrete function value domain space of behavior situation semantics and expand the state information entity; In step (1), six attributes are identified for each conversation, including conversation ID, content, behavior, object, status, and category.
2. The event knowledge graph GCN embedding representation method according to claim 1, characterized in that: The construction of GCN in step 3 includes: (1) Construct a vector file of word vector pairs based on the behavior type state space graph, and store the node and node pair information in the GCN graph; (2) Input the GCN construction algorithm and perform GCN embedding; (3) Output the behavioral basis of the labeled dataset dialogue.
3. An event knowledge graph GCN embedding representation system using the event knowledge graph GCN embedding representation method according to any one of claims 1 to 2, characterized in that: The event knowledge graph GCN embedding representation system includes: The data annotation module is used to predefine events and annotate the behavior state space attributes for each conversation data, and convert it into a behavior state space structured data set; The word vector model training module is used to define the behavior type for the training data set and the test data set, train the word vector model for the annotated content of the state space entity, and generate the word vector representing the annotated data set; The GCN construction module is used to construct a vector file of word vector pairs based on the behavior type state space graph, input the nodes and node pair information in the GCN graph to the GCN construction algorithm, and output the behavior basis of each dialogue data annotation dataset.
4. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the event knowledge graph GCN embedding representation method as described in any one of claims 1 to 2.
5. A computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the event knowledge graph GCN embedding representation method as described in any one of claims 1 to 2.
6. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the event knowledge graph GCN embedding representation system as described in claim 3.
Citation Information
Patent Citations
Dialogue behavior identification method and system based on conditional random field and structured attention network
CN108829662A
Domain event graph construction method and device fusing multiple types of facts and entity knowledge
CN110968700A
Event detection method based on upper concept information
CN113157859A
Event knowledge representation method based on behavior base
CN114004223A
Multi-modal event representation learning method based on event ontology
CN114201965A
