An event co-reference resolution method, device, equipment and readable storage medium

Abstract

The application provides an event co-reference resolution method, device and equipment and a readable storage medium. A graph structure data is constructed based on a dependency syntax relationship and event element information, and a mutual dependency of elements in a sentence is learned through a graph attention network. Context interaction information of an event pair of an event sentence is extracted through a PairCNN module, and the model can effectively solve the event co-reference resolution problem.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information extraction method, and particularly relates to an event coreference resolution method, device, equipment and readable storage medium. BACKGROUND

[0002] Event coreference resolution is a very important and challenging task in natural language processing, which aims to determine whether two event descriptions refer to the same event. Event coreference resolution is currently in the initial exploration stage, and there will inevitably be some shortcomings. The main shortcomings of most current event coreference resolution models are: (1) the relevance between event elements and the importance of the relevance are not considered; (2) the context interaction information of event pairs cannot be fully captured.

[0003] Therefore, the present application is proposed. SUMMARY

[0004] The present application discloses an event coreference resolution method, device, equipment and readable storage medium, which aims to solve the problems in the prior art.

[0005] The first embodiment of the present application provides an event coreference resolution method, comprising:

[0006] obtaining an event pair and constructing event pair graph structure data based on the event pair;

[0007] calling a graph attention network to perform a node encoding update operation to generate node information of a representative event of the graph structure data;

[0008] calling a PairCNN to learn event sentences in the event pair to generate context interaction information of the event sentences;

[0009] fusing the interaction information and the node information of the representative event to generate fusion information, and generating a coreference relationship of the event pair based on the fusion information.

[0010] Preferably, the event pair is obtained, and event pair graph structure data is constructed based on the event pair, specifically:

[0011] obtaining two event sentences of the event pair;

[0012] constructing two event sentence graph structure data respectively, and fusing the two event sentence graph structure data to generate event pair graph structure data;

[0013] Each of the event sentence graph structure data includes dependency graph structure data, event element graph structure data of a bidirectional connection edge between event element word sequences, self-connection graph structure data with the word itself as the starting point and the ending point, and event element type graph structure data with the mapping relationship between the event element type and the first and last words of the event element as the edge.

[0014] The event pair graph structure data is formed by connecting the event nodes with the corresponding event elements.

[0015] Preferably, the updating operation of calling the graph attention network for node encoding is used to generate node information of the representative event of the graph structure data, specifically:

[0016] The BERT pre-training model is called to perform vector encoding on the two event sentences to generate node initialization vectors of the graph structure data of the two event sentences, wherein the initialization vectors include word vectors, event element type vectors, and event node vectors.

[0017] The GAT is called to perform node embedding learning on the graph structure data and the node initialization vectors to generate updated vectors.

[0018] The attention coefficients between any two nodes in the graph structure data are generated according to the updated vectors.

[0019] The attention coefficients and the updated vectors are weighted and summed to generate node information of the representative event of the graph structure data.

[0020] Preferably, the PairCNN is called to learn the event sentences in the event pair to generate context interaction information of the event sentences, specifically:

[0021] The PairCNN is called to perform feature extraction on the word vectors of the event pair to generate an extraction result.

[0022] A similarity function is called to operate on the extraction result to generate similarity information.

[0023] A feedforward neural network is called to process the extraction result and the similarity information to generate context interaction information of the event sentences.

[0024] Preferably, the interaction information and the node information of the representative event are fused to generate fusion information, and the co-reference relationship of the event pair is generated based on the fusion information, specifically:

[0025] The event node vector and the interaction information are fused to generate a fusion vector.

[0026] A classifier is called to determine whether the event pair has a co-reference relationship.

[0027] Preferably, before the weighted sum according to the attention coefficient and the updated vector is generated, the node information of the representative event of the graph structure data, the method further comprises:

[0028] The attention coefficient is normalized.

[0029] The second embodiment of the application provides an event co-reference resolution device, comprising:

[0030] A graph structure data construction unit is configured to acquire an event pair and construct event pair graph structure data based on the event pair;

[0031] A node information generation unit is configured to call a graph attention network to perform a node encoding update operation and generate node information of a representative event of the graph structure data;

[0032] An interaction information generation unit is configured to call a PairCNN to learn event sentences in the event pair and generate context interaction information of the event sentences;

[0033] A co-reference relationship generation unit is configured to fuse the interaction information and the node information of the representative event to generate fusion information and generate a co-reference relationship of the event pair based on the fusion information.

[0034] The third embodiment of the application provides an event co-reference resolution device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the event co-reference resolution method according to any one of the above.

[0035] The fourth embodiment of the application provides a computer readable storage medium, characterized in that the computer readable storage medium stores a computer program, and the computer program can be executed by a processor of a device where the computer readable storage medium is located to implement the event co-reference resolution method according to any one of the above.

[0036] Based on the event co-reference resolution method, device, equipment and readable storage medium provided by the application, the graph structure data is constructed based on the dependency syntax relationship and the event element information, and the mutual dependency of the elements in the sentence is learned through the graph attention network. The context interaction information of the event sentences in the event pair is extracted through the PairCNN module, and the event co-reference resolution problem can be effectively solved. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a flowchart of the event co-reference resolution method provided by the first embodiment of the application;

[0038] Figure 2is a structural schematic diagram of an event co-reference resolution model provided by the present application.

[0039] Figure 3 is a module schematic diagram of an event co-reference resolution device provided by the second embodiment of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0041] In order to better understand the technical solutions of the present application, the embodiments of the present application will be described in detail below with reference to the drawings.

[0042] It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] The terms used in the embodiments of the present application are only for the purpose of describing the specific embodiments, and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0044] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0045] Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if determined" or "if detecting (a stated condition or event)" can be interpreted as "when determined" or "in response to determining" or "when detecting (a stated condition or event)" or "in response to detecting (a stated condition or event)".

[0046] The "first / second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order of the objects. Understandably, the "first / second" can be interchanged in a specific order or sequence as allowed. It should be understood that the objects distinguished by the "first / second" can be interchanged under appropriate circumstances to enable the embodiments described herein to be implemented in an order other than those illustrated or described herein.

[0047] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0048] The present application discloses an event co-reference resolution method, device, equipment and readable storage medium, aiming at solving the deficiencies in the prior art.

[0049] Please refer to Figure 1 The first embodiment of the present application provides an event co-reference resolution method, which can be executed by an event co-reference resolution device (hereinafter referred to as a resolution device), and in particular, by one or more processors in the resolution device, to at least implement the following steps:

[0050] S101, obtaining an event pair and constructing event pair graph structure data based on the event pair;

[0051] It should be noted that the resolution device can be a desktop computer, a notebook computer, a server, a workstation, or other terminal with data processing and analysis capabilities. The simulation device can be installed with a corresponding operating system and application software, and the functions required by the present embodiment can be realized through the combination of the operating system and the application software.

[0052] Specifically, in the present embodiment,

[0053] Obtaining two event sentences of the event pair;

[0054] Constructing two event sentence graph structure data respectively, and fusing the two event sentence graph structure data to generate event pair graph structure data;

[0055] Each of the event sentence graph structure data includes dependency graph structure data, event element graph structure data of a bidirectional connection edge between event element word sequences, self-connection graph structure data with the word itself as the starting point and the ending point, and event element type graph structure data with the mapping relationship between event element types and event element head and tail words as edges;

[0056] The event pair graph structure data is composed of event nodes and corresponding event elements connected by the event nodes.

[0057] More specifically:

[0058] Step 1.1, defining two event sentences of the input event pair as and in and Representing event sentence S respectively a and event sentence S b In the i-th and j-th characters, KA represents S. a The length, KB represents S b The length.

[0059] Step 1.2: According to Step 1.1, process the event statement S. a Dependency parsing is performed, and dependency graph structure data is constructed based on the parsing results. in KA1 = KA represents the set of nodes in the dependency graph structure data. The length of the node set, RA1 represents the set of dependency edges in a dependency graph structure. The length of the edge set. For each event element, construct the event element graph structure data, showing bidirectional connections between the event element word sequences. in The set of nodes representing the event element graph structure data, KA2 = KA represents The length of the node set, RA2 represents the set of bidirectional connections between event element word sequences in the event element graph structure data. The length of the edge set. By connecting characters themselves, a self-connected graph structure is constructed with each character as both the starting and ending point. represent The set of nodes, KA3 = KA represents The length of the node set, RA3 represents the set of edges in a self-connected graph structure. The length of the edge set. Construct an event element type graph structure data where event element types are nodes, and the mapping relationship between event element types and the first and last characters of event elements is the edge. in KE4 represents the set of nodes that represent the event element graph structure data. The length of the set of nodes in the middle, where Partial representation byte points in Partial representation Nodes of element type in In the graph structure data representing event element types, the mapping relationship between event element type nodes and the first and last characters of event elements is represented by a set of connecting edges. RA4 represents... Length of edge set. Fuse G1, G2, G3, G4, and merge the same edges to construct the fusion graph structure data G a =(E a ,R a ), wherein represents the nodes of the fusion graph structure data, KAT represents the number of words in the sentence text plus the number of event element type nodes, represents the edge set RAT of the fusion graph structure data G a Length of edge set.

[0060] The same operation is performed on the event sentence S b (not described here), and the fusion graph structure data G b of S b =(E b ,R b ) is obtained, and then G a is merged with G b , and only one copy of the same event element is retained. At the same time, two nodes representing events, called event nodes, are added, and the event nodes are connected with the respective event elements to construct the event pair graph structure data G pair =(E pair ,R pair ), wherein represents the set of nodes of G pair , PE represents the length of the node set of G pair , represents the set of edges of G pair , PR represents the length of the edge set of G pair .

[0061] S102, calling the graph attention network to perform the updating operation of node encoding to generate the node information of the event representing node of the graph structure data;

[0062] Specifically, in the embodiment:

[0063] The BERT pre-training model is called to perform vector encoding on the two event sentences to generate the node initialization vectors of the graph structure data of the two event sentences, wherein the initialization vectors include word vectors, event element type vectors, and event node vectors;

[0064] The GAT is called to perform node embedding learning on the graph structure data and the node initialization vectors to generate updated vectors;

[0065] The attention coefficients between any two nodes in the graph structure data are generated according to the updated vectors;

[0066] The node information of a representative event of the graph structure data is generated by weighted summation according to the attention coefficient and the updated vector.

[0067] More specifically:

[0068] Step 2.1, vector encoding of any word in S a and S b is obtained to obtain the corresponding feature vector and where AE, BE respectively represent the text length of S a and S b , d0 represents the dimension of the vector. Then, the event element type is vector encoded by BERT to obtain the feature vector corresponding to the event element type

[0069] The event node vector in the graph structure data is initialized as a 0 vector with a dimension of d0. The word vector, event element type vector and event node vector of S a and S b are used to initialize the vector representation of the nodes of G pair , and the node vector set E pair of G pair is obtained. where PE is the number of nodes of G pair .

[0070] Step 2.2, the constructed graph structure data G pair and the vector V E corresponding to the nodes of G pair are input into GAT for node embedding learning to obtain the updated vector where PEG=PE is the number of nodes of G pair .

[0071] Step 2.3, the calculation of the attention coefficient e ij between any two nodes in the graph structure data is performed, as shown in formula (1):

[0072]

[0073] where e ij represents the attention coefficient between two nodes, represents the vector of the i-th node in V Eg , represents the vector of the j-th node in V Eg , represents the neighbor set of the i-th node, denotes concatenating two vectors, o(·) denotes a single-layer feedforward neural network o(d1×d1)→R; the attention coefficient normalization process is shown in equation (2):

[0074]

[0075] where k represents the kth neighbor node of the ith node, exp represents the take pointer function, LeakyReLU represents the nonlinear activation function, and the value of the function score less than 0 is set to -0.2, as shown in equation (3):

[0076]

[0077] According to the calculation result, the vector encoding of node i is weighted and summed, as shown in equation (4):

[0078]

[0079] where, represents the learned vector encoding, σ represents the sigmoid function, and α ij represents the attention coefficient between two nodes, represents the vector encoding of the neighbor node j of the node. The vector encoding of the event pair graph structure data obtained after learning by the graph attention network is

[0080] S103, calling PairCNN to learn the event sentences in the event pair to generate context interaction information of the event sentences;

[0081] Specifically, in the present embodiment:

[0082] calling PairCNN to extract features of the word vectors of the event pair to generate an extraction result;

[0083] calling a similarity function to operate on the extraction result to generate similarity information;

[0084] calling a feedforward neural network to process the extraction result and the similarity information to generate context interaction information of the event sentences.

[0085] More specifically:

[0086] Step 3.1 inputs the vector representation V AE and V BE of the event sentences in the event pair into PairCNN to learn the context interaction information of the event sentences in the event pair. Specifically, first, V AE and V BEThe input into the CNN obtains the vector representation of the two event sentences, as shown in the following formulas (5) and (6):

[0087]

[0088]

[0089] wherein, CNN represents a convolutional neural network, V AE and V BE represent the vector representation of the event sentences in the event pair, represent the features extracted by the CNN, and be input into a similarity function for calculation, as shown in the following formula (7):

[0090]

[0091] wherein, represents a similarity matrix, and sin represents a similarity calculation function. The and input into a feedforward neural network, as shown in the following formula (8):

[0092]

[0093] wherein, represents the context interaction information of the event sentences in the event pair, and feedforward represents a feedforward neural network.

[0094] S104, fuse the interaction information and the node information representing the event to generate fusion information, and generate the coreference relationship of the event pair based on the fusion information.

[0095] Specifically, in the present embodiment:

[0096] The fusion of the interaction information and the node information representing the event to generate fusion information, and the generation of the coreference relationship of the event pair based on the fusion information are specifically:

[0097] fuse the event node vector and the interaction information to generate a fusion vector;

[0098] call a classifier to judge whether the fusion vector has a coreference relationship.

[0099] More specifically: step 4.1 concatenates the vector pair representing the event node in G and the context interaction information vector to obtain an event pair fusion vector, as shown in the following formula (9):

[0100]

[0101] wherein, and represents the graph structure data G pair of the event nodes in the event pair. Finally, the fusion vector is sent into the classifier to determine whether the event pair has a coreference relationship.

[0102] Step 4.2, according to the above operation, the loss function of the event coreference resolution model (as shown in Figure 2 ) that integrates the interaction information of the event pair is as follows (10):

[0103]

[0104] wherein, C represents the number of classifications, which is set to 2 here, x, y represent the input event pair, l represents the true label of whether the event pair is a coreference relationship, represents the predicted label of whether the event pair is a coreference relationship, and log represents the logarithmic function. In addition, in order to prevent overfitting, L2-regularization is also introduced. The final training target is as shown in the following formula (11):

[0105]

[0106] wherein, θ represents the learning parameter, λ represents the parameter of L2-regularization, n represents the total number of samples, r represents the rth training sample, represents the predicted label of whether the rth sample has a coreference relationship, and l (r) represents the true label of whether the rth sample has a coreference relationship.

[0107] The following uses a simple example to further illustrate the working process of the entire coreference resolution model:

[0108] The input test event data example is as follows, e1 represents one of the event descriptions in the event pair, e2 represents the other event description in the event pair, text represents the event sentence, trigger represents the trigger word and the event type corresponding to the trigger word, argument represents the argument, and output represents the label of whether the event pair is coreferential.

[0109] [e1: {"text": "PTN chain end station cross board hidden problem causes part of the TDS site PS down rate index deterioration and interactive business abnormal problem", "trigger": ["IndexFault", 31, "deterioration"], "argument": [[ "Index", 18, "TDS site PS down rate index "]]},

[0110] e2: {"text": "PS drop rate increases", "trigger": ["IndexFault", 5, "increase"], "argument": [[ "Index", 0, "PS drop rate"]]}}

[0111] Output: True

[0112] See Figure 3 The second embodiment of the application provides an event co-reference resolution device, comprising:

[0113] A graph structure data construction unit 201 is configured to acquire event pairs and construct event pair graph structure data based on the event pairs.

[0114] A node information generation unit 202 is configured to call a graph attention network to perform a node encoding update operation and generate node information of representative events of the graph structure data.

[0115] An interaction information generation unit 203 is configured to call a PairCNN to learn event sentences in the event pairs to generate context interaction information of the event sentences.

[0116] A co-reference relationship generation unit 204 is configured to fuse the interaction information and the node information of the representative events to generate fusion information and generate co-reference relationships of the event pairs based on the fusion information.

[0117] The third embodiment of the application provides an event co-reference resolution device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the event co-reference resolution method according to any one of the above embodiments.

[0118] The fourth embodiment of the application provides a computer readable storage medium, characterized in that the computer readable storage medium stores a computer program, and the computer program can be executed by a processor of a device where the computer readable storage medium is located to implement the event co-reference resolution method according to any one of the above embodiments.

[0119] The application provides an event co-reference resolution method, device, equipment and readable storage medium. The graph structure data is constructed based on the dependency syntax relationship and event element information, and the mutual dependency of elements in a sentence is learned through a graph attention network. The context interaction information of event sentences in event pairs is extracted through a PairCNN module, and the event co-reference resolution problem can be effectively solved.

[0120] Exemplarily, the computer program in the third and fourth embodiments of the present application can be divided into one or more modules stored in the memory and executed by the processor to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the event coreference resolution device. For example, the device in the second embodiment of the present application.

[0121] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like, which is a control center of the event coreference resolution method and is connected to each part of the event coreference resolution method through various interfaces and lines.

[0122] The memory can be used to store the computer program and / or the module, and the processor realizes various functions of the event coreference resolution method by running or executing the computer program and / or the module stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, a character conversion function, etc.), etc.; and the data storage area can store data created according to the use of the mobile phone (such as audio data, text message data, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a nonvolatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0123] The implemented modules, if implemented in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods of the present application can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the contents included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0124] It should be noted that the above-described device embodiments are only schematic, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. In addition, the connection relationship between the modules in the device embodiment provided by the present application indicates that there is a communication connection between them, which can be realized as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0125] The above is only a preferred embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of resolution of event coreference, characterized by, The method comprises the following steps: obtaining an event pair and constructing event pair graph structure data based on the event pair; calling a graph attention network to perform an updating operation of node coding to generate node information of a representative event of the graph structure data; calling a PairCNN to learn event sentences in the event pair to generate context interaction information of the event sentences; fusing the interaction information and the node information of the representative event to generate fusion information, and generating a coreference relationship of the event pair based on the fusion information; wherein the step of obtaining an event pair and constructing event pair graph structure data comprises the following steps: obtaining two event sentences of the event pair; constructing two event sentence graph structure data respectively, and fusing the two event sentence graph structure data to generate event pair graph structure data; wherein each of the event sentence graph structure data comprises dependency graph structure data, event element graph structure data of a bidirectional connection edge between event element word sequences, self-connection graph structure data with a word as a starting point and an ending point, and event element type graph structure data with a mapping relationship between event element types and head and tail words as an edge; wherein the event pair graph structure data is formed by connecting event nodes and corresponding event elements; the step of calling a graph attention network to perform an updating operation of node coding to generate node information of a representative event of the graph structure data comprises the following steps: calling a BERT pre-training model to perform vector coding on the two event sentences to generate node initialization vectors of the graph structure data of the two event sentences, wherein the initialization vectors comprise word vectors, event element type vectors and event node vectors; calling a GAT to perform node embedding learning on the graph structure data and the node initialization vectors to generate updated vectors; generating attention coefficients between any two nodes in the graph structure data according to the updated vectors; performing weighted summation on the attention coefficients and the updated vectors to generate node information of a representative event of the graph structure data; the step of calling a PairCNN to learn event sentences in the event pair to generate context interaction information of the event sentences comprises the following steps: calling a PairCNN to perform feature extraction on word vectors of the event pair to generate extraction results; calling a similarity function to operate on the extraction results to generate similarity information; calling a feedforward neural network to process the extraction results and the similarity information to generate context interaction information of the event sentences; the step of fusing the interaction information and the node information of the representative event to generate fusion information, and generating a coreference relationship of the event pair based on the fusion information comprises the following steps: fusing the event node vectors and the interaction information to generate fusion vectors; calling a classifier to judge whether the event pair has a coreference relationship.

2. The method of resolving an event coreference of claim 1, wherein, Before the step of performing weighted summation on the attention coefficients and the updated vectors to generate node information of a representative event of the graph structure data, the method further comprises the following step: normalizing the attention coefficients.

3. An apparatus for resolution of event coreference, the apparatus comprising: The method comprises the following steps: The graph structure data construction unit is configured to obtain an event pair and construct event pair graph structure data based on the event pair. The node information generation unit is configured to call a graph attention network to perform a node encoding update operation and generate node information of a representative event of the graph structure data. The interaction information generation unit is configured to call a PairCNN to learn event sentences in the event pair and generate context interaction information of the event sentences. The coreference relation generation unit is configured to fuse the interaction information and the node information of the representative event to generate fusion information, and generate a coreference relation of the event pair based on the fusion information. The graph structure data construction unit is configured to obtain two event sentences of an event pair, construct two event sentence graph structure data respectively, and fuse the two event sentence graph structure data to generate event pair graph structure data. Each of the event sentence graph structure data includes dependency graph structure data, event element graph structure data of a bidirectional connection edge between event element word sequences, self-connection graph structure data with a word as a starting point and an ending point, and event element type graph structure data with a mapping relationship between event element types and head / tail words as edges. The node information generation unit is configured to call a BERT pre-training model to perform vector encoding on the two event sentences to generate node initialization vectors of the graph structure data of the two event sentences, wherein the initialization vectors include word vectors, event element type vectors, and event node vectors. The node information generation unit is configured to call a GAT to perform node embedding learning on the graph structure data and the node initialization vectors to generate updated vectors, generate attention coefficients between any two nodes in the graph structure data according to the updated vectors, and perform weighted summation on the attention coefficients and the updated vectors to generate node information of a representative event of the graph structure data. The interaction information generation unit is configured to call a PairCNN to perform feature extraction on word vectors of the event pair to generate extraction results, call a similarity function to operate on the extraction results to generate similarity information, and call a feedforward neural network to process the extraction results and the similarity information to generate context interaction information of the event sentences. The coreference relation generation unit is configured to fuse the event node vectors and the interaction information to generate fusion vectors, and call a classifier to determine whether the event pair has a coreference relation.

4. An event coreference resolution apparatus, characterized by The coreference relation generation unit is configured to fuse the event node vectors and the interaction information to generate fusion vectors, and call a classifier to determine whether the event pair has a coreference relation.

5. A computer readable storage medium, characterized in that, The computer program can be executed by the processor to implement the event coreference resolution method of any one of claims 1-2. The computer program can be executed by the processor to implement the event coreference resolution method of any one of claims 1-2.

Images