An event detection method and device based on a multi-layer graph attention network

Abstract

The application provides an event detection method and device based on a multi-layer graph attention network, comprising obtaining context words in event text information, and determining a syntactic information adjacency matrix and a concatenation vector corresponding to the context words; taking the adjacency matrix and the concatenation vector as inputs of an artificial neural network to obtain an output vector; generating aggregated information according to the concatenation vector and the output vector; and determining a trigger word category of the context words according to the aggregated information. The application effectively solves the problem of information loss and error propagation that easily occurs when a syntactic analysis tool is used by simultaneously combining syntactic information and context information of the context words; and avoids the situation that the classification of the final trigger word is not ideal due to the excessive propagation of some short-distance syntactic information by combining a skip connection module in the graph attention network layer, thereby effectively improving the precision, recall rate and F1 value of trigger word classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing, and in particular to an event detection method and device based on a multi-layer graph attention network. BACKGROUND

[0002] Knowledge Graph describes the concepts, entities and their relationships in the objective world in a structured form, expresses the information of the Internet in a form closer to human cognition of the world, and provides the ability to better organize, manage and understand the massive information of the Internet. Knowledge Graph was proposed by Google in 2012 and successfully applied to search engines. Knowledge Graph belongs to the research field of artificial intelligence, namely knowledge engineering, and is a killer application of large-scale knowledge resources established by knowledge engineering. Typical examples are Google's acquisition of Freebase (a free knowledge database) and the launch of Knowledge Graph in 2012, Facebook's graph search, Microsoft Satori, and specific knowledge bases in the fields of business, finance, and life sciences.

[0003] Event knowledge in the Knowledge Graph is implicit in Internet resources, including existing structured semantic knowledge, structured information of databases, semi-structured information resources, and unstructured resources. Different resources have different knowledge acquisition methods. The recognition and extraction of events researches how to recognize and extract event information from text describing event information and present it in a structured form, including the time, place, participating roles, and related actions or changes in state.

[0004] Traditional event detection methods ignore the syntactic features contained in the words in the sentence, only use sentence-level features, and event detection is prone to low recognition efficiency and classification accuracy of trigger words due to word ambiguity. In recent years, methods using syntactic information to improve event detection have been proven to be very effective. For example, the paper "Non-trigger event detection method based on syntactic information fusion" proposes to use syntactic information combined with attention mechanism (ATTENTION) to realize the connection of dispersed event information in the sentence to improve the accuracy of event detection; the paper "Vietnamese news event detection based on dependency information and convolutional neural network" uses convolutional coding to fuse the features between non-continuous words with dependency syntactic information, and then fuses the two parts of the features as event coding to realize event detection. SUMMARY

[0005] In view of the problems, the present application is proposed to provide an event detection method based on a multi-layer graph attention network to overcome the problems or at least partially solve the problems, which comprises the steps of:

[0006] An event detection method based on a multi-layer graph attention network, comprising the steps of:

[0007] Obtaining context words in event text information, and determining a syntactic information adjacency matrix and a concatenation vector corresponding to the context words;

[0008] Taking the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and obtaining an output vector;

[0009] Generating aggregated information according to the concatenation vector and the output vector;

[0010] Determining a trigger word class of the context words according to the aggregated information.

[0011] Further, the step of obtaining context words in event text information, and determining a syntactic information adjacency matrix and a concatenation vector corresponding to the context words, comprises:

[0012] Determining syntactic information corresponding to the context words according to the context words;

[0013] Generating the syntactic information adjacency matrix according to the syntactic information;

[0014] Generating the concatenation vector according to a word embedding vector of the context words.

[0015] Further, the step of determining syntactic information corresponding to the context words according to the context words, comprises:

[0016] Analyzing the event text information through syntactic dependency, and generating the syntactic information corresponding to the context words according to an analysis result of the event text information.

[0017] Further, the step of taking the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and obtaining an output vector, comprises:

[0018] Generating a tensor from the adjacency matrix of the same batch;

[0019] Inputting the tensor and the concatenation vector into an artificial neural network for calculation, and generating the output vector according to a calculation result of the artificial neural network.

[0020] Further, the step of determining a trigger word class of the context words according to the aggregated information, comprises:

[0021] Determining a trigger word of the context words according to the aggregated information, and classifying the trigger word according to a classifier module.

[0022] An event detection device based on a multi-layer graph attention network, comprising:

[0023] An acquisition module is configured to acquire a context word in event text information and determine a syntactic information adjacency matrix and a concatenation vector corresponding to the context word;

[0024] A calculation module is configured to take the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and acquire an output vector;

[0025] An aggregation module is configured to aggregate the concatenation vector and the output vector to generate aggregated information;

[0026] A classification module is configured to determine a trigger word category of the context word according to the aggregated information.

[0027] Further, the acquisition module comprises:

[0028] An expression submodule is configured to determine syntactic information corresponding to the context word according to the context word;

[0029] A generation submodule is configured to generate the syntactic information adjacency matrix according to the syntactic information;

[0030] A concatenation submodule is configured to generate the concatenation vector according to a word embedding vector of the context word.

[0031] Further, the expression submodule comprises:

[0032] A dependency analysis submodule is configured to analyze the event text information through syntactic dependency, and generate the syntactic information corresponding to the context word according to an analysis result of the event text information.

[0033] Further, the calculation module comprises:

[0034] An array conversion submodule is configured to generate a tensor from the adjacency matrices of the same batch;

[0035] An artificial neural network calculation submodule is configured to input the tensor and the concatenation vector into an artificial neural network for calculation, and generate the output vector according to a calculation result of the artificial neural network.

[0036] Further, the classification module comprises:

[0037] A trigger word processing submodule is configured to determine a trigger word of the context word according to the aggregated information, and classify the trigger word according to a classifier module.

[0038] The present application has the following advantages:

[0039] In the embodiments of the present application, by acquiring a context word in event text information, syntactic information adjacency matrix and concatenation vector corresponding to the context word are determined; the adjacency matrix and the concatenation vector are taken as inputs of an artificial neural network to acquire an output vector; aggregated information is generated according to the concatenation vector and the output vector; and a trigger word category of the context word is determined according to the aggregated information. By simultaneously combining syntactic information and context information of the context word, the present application can effectively solve the problems of information loss and error propagation that easily occur when a syntactic analysis tool is used; and by combining a skip connection module in a graph attention network layer, more original features can be retained to avoid the situation that the classification of the final trigger word is not ideal due to the excessive propagation of some short-distance syntactic information, thereby effectively improving the precision, recall rate and F1 value of trigger word classification. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the present application, the drawings needed to be used in the description of the present application will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0041] Figure 1 is a step flowchart of an event detection method based on a multi-layer graph attention network provided by an embodiment of the present application;

[0042] Figure 2 is a schematic diagram of a syntactic dependency tree provided by an embodiment of the present application;

[0043] Figure 3 is a schematic diagram of an adjacency matrix provided by an embodiment of the present application;

[0044] Figure 4 is a schematic diagram of a graph attention network provided by an embodiment of the present application;

[0045] Figure 5 is a flowchart of an event detection method based on a multi-layer graph attention network provided by an embodiment of the present application;

[0046] Figure 6 is a structural block diagram of an event detection device based on a multi-layer graph attention network provided by an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0048] With reference to Figure 1 , an event detection method based on a multi-layer graph attention network is shown, which is provided by an embodiment of the present application;

[0049] The method comprises:

[0050] S110, obtaining context words in event text information, and determining a syntactic information adjacency matrix and a concatenation vector corresponding to the context words;

[0051] S120, taking the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and obtaining an output vector;

[0052] S130, generating aggregated information according to aggregation of the concatenation vector and the output vector;

[0053] S140, determining a trigger word category of the context words according to the aggregated information.

[0054] In the embodiments of the present application, by obtaining context words in event text information, and determining a syntactic information adjacency matrix and a concatenation vector corresponding to the context words; taking the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and obtaining an output vector; generating aggregated information according to aggregation of the concatenation vector and the output vector; and determining a trigger word category of the context words according to the aggregated information. By simultaneously combining syntactic information and context information of the context words, the present application can effectively solve the problem of information loss and error propagation that easily occurs when using a syntactic analysis tool; and by combining a skip connection module in a graph attention network layer, more original features can be retained, and the situation of unsatisfactory classification of the final trigger word due to excessive propagation of some short-distance syntactic information can be avoided, thereby effectively improving the precision, recall rate and F1 value of trigger word classification.

[0055] In the following, the event detection method based on a multi-layer graph attention network in the present exemplary embodiment will be further described.

[0056] As described in step S110, context words in event text information are obtained, and a syntactic information adjacency matrix and a concatenation vector corresponding to the context words are determined.

[0057] In an embodiment of the present application, the specific process of "obtaining the context words in the event text information and determining the syntactic information adjacency matrix and concatenation vector corresponding to the context words" in step S110 can be further illustrated in combination with the following description.

[0058] The syntactic information corresponding to the context words is determined according to the context words, as described in the following steps.

[0059] In an embodiment of the present application, the specific process of "determining the syntactic information corresponding to the context words according to the context words" can be further illustrated in combination with the following description.

[0060] The event text information is analyzed by syntactic dependency, and the syntactic information corresponding to the context words is generated according to the analysis result of the event text information, as described in the following steps.

[0061] It should be noted that the syntactic dependency is to reveal the syntactic structure of the language unit by analyzing the dependency relationship between the components, and the syntactic dependency analysis identifies the "subject-predicate-object" and "determining-complement" in the sentence, and emphasizes the relationship between the words. The core of the sentence in the syntactic dependency analysis is the predicate verb, and then other components are found around the predicate, and finally the sentence is analyzed into a syntactic dependency tree. The syntactic dependency tree can describe the dependency relationship between each word.

[0062] In a specific implementation, the event text information is obtained, the event text information is recognized, the StanfordCore NLP (StandFord Natural Language Processing) is used for syntactic dependency analysis, each sentence in the event text is analyzed, the event trigger word in the sentence is recognized, and the dependency relationship between the event trigger word and the event argument, and / or the event argument and the event argument is emphasized. Analysis, form a syntactic dependency tree.

[0063] Among them, the event trigger word refers to a word that best represents the occurrence of an event in an event, which is the projection of the event concept at the word and phrase level, and is the basis and means of event recognition. It is also an important feature that determines the event category, and is generally a verb or noun; the event argument refers to information describing the time, place and person of the event occurrence.

[0064] Referring to Figure 2The figure shows a schematic diagram of a syntactic dependency tree provided in an embodiment of this application. As shown in the figure, in the constructed syntactic dependency tree of the sentence "I went to Tiananmen Square in Beijing to watch the sunrise", the core predicate of the sentence is "go", which is the root of the syntactic dependency tree. The subject of "go" is "I", the object of "go" is "Tiananmen Square in Beijing", and the object of the other verb "watch" is "sun". The syntactic dependency tree can describe the dependency relationship between the context words.

[0065] As described in the following steps, the syntactic information adjacency matrix is ​​generated based on the syntactic information;

[0066] It should be noted that the adjacency matrix is ​​a matrix representing the adjacency relationships between vertices. Let G = (V, E) be a graph, where V = {v1, v2, ..., v...} n Let V be a vertex and E be an edge. A one-dimensional array stores all vertex data in the graph, and a two-dimensional array stores the relationships (edges or arcs) between vertices. This two-dimensional array is called the adjacency matrix. Adjacency matrices are further divided into directed graph adjacency matrices and undirected graph adjacency matrices. The adjacency matrix of G is an n-order square matrix with the following properties: For undirected graphs, the adjacency matrix is ​​always symmetric, and the main diagonal is always zero, while the secondary diagonal is not necessarily zero; this is not necessarily true for directed graphs. In an undirected graph, the degree of any vertex i is the number of all non-zero elements in the i-th column (or i-th row). In a directed graph, the out-degree of vertex i is the number of all non-zero elements in the i-th row, and the in-degree is the number of all non-zero elements in the i-th column. The adjacency matrix of a directed graph is used to store the syntactic dependencies between two event parameters.

[0067] As an example, each sentence is processed through syntactic dependency parsing to form a syntactic dependency tree, and then the corresponding adjacency matrix is ​​generated based on the syntactic dependency tree.

[0068] In a specific implementation, refer to Figure 3 The diagram shows an adjacency matrix provided in one embodiment of this application. Figure 3 The adjacency matrix shown corresponds to Figure 2 The syntactic dependency tree shown. Figure 2The trigger words in the sentence are "go", "Beijing", and "Tiananmen", which are parallel objects. Therefore, in the corresponding adjacency matrix, the intersection of the row where "go" is located and the column where "Beijing" and "Tiananmen" are located has a value of 1. Each word is a node, and "I", "go", "Beijing", "Tiananmen", "see", "sun", and "rise" are seven words, so a 7x7 matrix is formed. If there is a syntactic arc between two words, the corresponding position of the matrix is 1, otherwise it is 0. The adjacency matrix of the directed graph is used to store the syntactic dependency relationship of the text. If there is a dependency relationship between words, the corresponding adjacency matrix element value is 1, and if there is no dependency relationship between words, the corresponding adjacency matrix element is 0. The dependency relationship between the context words can be represented by the adjacency matrix.

[0069] The concatenation vector is generated according to the word embedding vector of the context word, as described in the following steps.

[0070] It should be noted that the word-level information in the sentence needs to be converted into a real-valued vector as the input of the artificial neural network. Let X = {x1, x2, x3, …, xn} be a sentence of length n, where xi is the i-th word in the sentence. In natural language processing tasks, the semantic information of a word is related to its position in the sentence, and the part-of-speech and entity type information can improve the recognition of trigger words and the understanding of semantics. The concatenation vector obtained by concatenating the word meaning vector, entity vector, part-of-speech vector, and position vector of the context word is used as the input of the artificial neural network.

[0071] In a specific implementation, the word meaning vector, entity vector, part-of-speech vector, and position vector of the context word are concatenated into a first concatenation vector, and the first concatenation vector is input into a Bi-LSTM neural network layer to generate a second concatenation vector. The second concatenation vector is one of the input vectors of the multi-layer graph attention network, and the concatenation vector can obtain the semantic information between the context words.

[0072] As described in step S120, the adjacency matrix and the concatenation vector are used as the input of the artificial neural network to obtain an output vector.

[0073] It should be noted that the artificial neural network is a graph attention network. Since the traditional graph convolution network has various limitations, it cannot well process directed graphs, cannot be applied to inductive tasks (inductive tasks refer to different graph structures that need to be processed in the training stage and the test stage), and cannot process dynamic graphs, while the graph attention network can well solve the defects of the graph convolution network. For each node, the similarity coefficient of node j to node i can be calculated using the attention mechanism, so that the graph structure can be used without relying entirely on the graph structure, and the inductive task can be applied. Under the graph attention network, even if the structure of the graph is changed during prediction, the influence on the graph attention network is small, and only the parameters need to be adjusted and recalculated. The operation mode of the graph attention network is vertex-by-vertex operation, and all vertices on the graph need to be looped through to complete each operation. Vertex-by-vertex operation means that the graph is freed from the constraints of the Laplacian matrix in the original graph structure, so that the directed graph problem can be solved.

[0074] In an embodiment of the present application, the specific process of "taking the adjacency matrix and the concatenation vector as the input of the artificial neural network to obtain the output vector" in step S120 can be further described in combination with the following description.

[0075] The adjacency matrix of the same batch is formed into a tensor as described in the following steps.

[0076] In a specific implementation, the sentences identified at the same time in the event text information are a batch, the adjacency matrix of the sentences in the same batch is formed into a tensor, and the set of adjacency matrices is represented as The tensor is formed as A∈R N*N*K , where K=|T V |, and N is the number of nodes.

[0077] The tensor and the concatenation vector are input into the artificial neural network for calculation, and the output vector is generated according to the calculation result of the artificial neural network, as described in the following steps.

[0078] As an example, referring to Figure 4 , a schematic diagram of a graph attention network provided by an embodiment of the present application is shown, which is divided into two steps of calculating attention coefficients and weighted summation. The tensor and the second concatenation vector are taken as the input of the graph attention layer, represented as where N is the number of nodes, F is the number of node features; and the output is where F' represents the dimension of the new node feature vector. The attention coefficient of node i and the surrounding neighbor node j∈Ni is calculated as Figure 4The calculation formula is shown as follows on the left side:

[0079]

[0080] wherein a is an R F′ x R F′ mapping, W is an R F′×F weight matrix.

[0081] The graph attention network can use an attention mechanism to calculate the similarity coefficient weight of node i and neighbor node j for each node, so that the graph structure does not need to be completely relied on.

[0082] The attention coefficient is normalized by using softmax, and the calculation formula is shown as follows:

[0083]

[0084] wherein || represents vector splicing, e ij and α ij are called "attention coefficients", and α ij is normalized on the basis of e ij .

[0085] After the attention coefficients of all nodes are normalized, the features of the neighbor nodes are weighted and summed to generate an output vector, and the calculation formula is shown as follows:

[0086]

[0087] wherein W is a weight matrix multiplied with the features, σ is a nonlinear activation function, and j is an element of N i , which represents all nodes adjacent to i.

[0088] As shown on the right side, Figure 4 a three-layer graph attention network, a multi-layer attention mechanism allocates different attention weights to different features. For the multi-layer graph attention network, the calculation formula is as follows:

[0089]

[0090] If the multi-layer graph attention network is applied to the output layer, the calculation formula is shown as follows:

[0091]

[0092] As described in step S130, the aggregation information is generated according to the spliced vector and the output vector.

[0093] As an example, in the graph attention network of each layer, the aggregation of the syntax information is realized by a skip-connection module, the splicing vector is jumped through the skip-connection module through the graph attention network of each layer, and is aggregated with the output vector. The skip-connection module can prevent the over-propagation of short-distance syntax information, can retain more original syntax information, and can avoid the situation that the final trigger word classification effect is poor.

[0094] As described in step S140, the trigger word category of the context word is determined according to the aggregated information.

[0095] In an embodiment of the present application, the specific process of "determining the trigger word category of the context word according to the aggregated information" in step S140 can be further described in combination with the following description.

[0096] As described in the following steps, the trigger word of the context word is determined according to the aggregated information, and the trigger word is classified by the classifier module.

[0097] As an example, the trigger word of the context word is determined according to the aggregated information, the trigger word is classified by the preset condition of the classifier module, and the event type corresponding to the event sentence is determined according to the classification category of the trigger word. The event type is a predefined different type.

[0098] Specifically, the preset condition of the classifier module is to aggregate the information of different modules, pass through a fully connected layer, and then select the maximum one of the category probabilities corresponding to each context word as the label of the current trigger word prediction through a softmax function (the softmax function maps the outputs of multiple neurons to the interval (0, 1), which can be understood as a probability, so as to perform multi-classification).

[0099] Next, an experiment is performed to demonstrate the event detection method based on the multi-layer graph attention network proposed in the embodiments of the present application:

[0100] Experimental environment: Pytorch-1.8.0 (open source Python machine learning library), Nvidia GeForce RTX3060 (graphics card chip), Windows 10 (computer operating system), Inter i7-11700k, memory 16G, hard disk 1T.

[0101] The experimental data is shown in Table 1:

[0102] Table 1 Comparison results of experiments

[0103]

[0104] Experimental results: The experiment takes precision (P), recall (R), and F1-score as observation variables. The definitions of P, R, and F1-score are as follows:

[0105]

[0106]

[0107]

[0108] In order to ensure the accuracy of the experiment, the division of the data set in this experiment is consistent with the division of the data set of other event detection methods. The experimental results prove that the event detection method proposed in this embodiment is about 8% higher than the traditional event detection method which only uses sentence-level features in F1-score; compared with the method based on graph neural network, the event method proposed in this embodiment also achieves the highest value in F1-score and Recall.

[0109] Referring to Figure 5 , a flowchart of an event detection method based on a multi-layer graph attention network is shown;

[0110] In a specific implementation, after obtaining event text information, the event text information is analyzed by syntax analysis technology and a syntax dependency tree is generated. Then, an adjacency matrix corresponding to the context word is generated according to the syntax dependency tree, and a tensor is generated from the adjacency matrices of the same batch of sentences. Four different word embedding vectors of the context word are spliced into a first spliced vector, and the first spliced vector is input into a Bi-LSTM neural network layer to generate a second spliced vector. The adjacency matrix and the second spliced vector are input into a multi-layer graph attention network to generate an output vector, so as to perform aggregation operation on syntax information of different depths. The spliced vector is further aggregated by a skip connection module to skip the multi-layer graph attention network. The output vector and the spliced vector are aggregated, and a classifier module is used to classify the context word to determine the event type corresponding to the event sentence.

[0111] For the device embodiment, it is basically similar to the method embodiment, so the description is relatively simple. For related parts, refer to the part of the method embodiment.

[0112] Referring to Figure 6 , an event detection device based on a multi-layer graph attention network is shown;

[0113] Specifically, it includes:

[0114] The acquisition module 610 is configured to acquire a context word in event text information, and determine a syntactic information adjacency matrix and a concatenation vector corresponding to the context word.

[0115] The calculation module 620 is configured to take the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and acquire an output vector.

[0116] The aggregation module 630 is configured to aggregate the concatenation vector and the output vector to generate aggregated information.

[0117] The classification module 640 is configured to determine a trigger word category of the context word according to the aggregated information.

[0118] In an embodiment of the present application, the acquisition module 610 comprises:

[0119] The expression submodule is configured to determine syntactic information corresponding to the context word according to the context word.

[0120] The generation submodule is configured to generate the syntactic information adjacency matrix according to the syntactic information.

[0121] The concatenation submodule is configured to generate the concatenation vector according to a word embedding vector of the context word.

[0122] In an embodiment of the present application, the expression submodule comprises:

[0123] The dependency analysis submodule is configured to analyze the event text information through syntactic dependency, and generate the syntactic information corresponding to the context word according to an analysis result of the event text information.

[0124] In an embodiment of the present application, the calculation module 620 comprises:

[0125] The array conversion submodule is configured to generate a tensor from the adjacency matrices of the same batch.

[0126] The artificial neural network calculation submodule is configured to input the tensor and the concatenation vector into an artificial neural network for calculation, and generate the output vector according to a calculation result of the artificial neural network.

[0127] In an embodiment of the present application, the classification module 640 comprises:

[0128] The trigger word processing submodule is configured to determine a trigger word of the context word according to the aggregated information, and classify the trigger word according to a classifier module.

[0129] While the preferred embodiments of the application have been described above, it will be recognized and understood that various modifications and changes can be made to the embodiments of the present application by those skilled in the art having the benefit of this detailed description without departing from the scope of the present application. It is therefore intended that the appended claims be construed to include all such embodiments and modifications as fall within the scope of the present application.

[0130] Finally, it should be noted that, in the specification, relational terms such as first and second, and the like, can be used solely to distinguish one entity or action from another entity or action without necessarily implying any actual relationship or order between such entities or actions. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0131] The above describes in detail the event detection method and device based on the multi-layer graph attention network provided by the present application. The principles and implementation manners of the present application are described by using specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges can be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A multi-layer graph attention network-based event detection method, characterized in that, The method comprises the steps of: obtaining context words in event text information, and determining a syntactic information adjacency matrix and a concatenation vector corresponding to the context words, including that each sentence in the event text information forms a syntactic dependency tree through syntactic dependency analysis, and then a corresponding adjacency matrix is generated according to the syntactic dependency tree; the concatenation vector is generated according to a word embedding vector of the context words; four different word embedding vectors of a word sense vector, an entity vector, a part-of-speech vector and a position vector of the context words are concatenated into a first concatenation vector, and then the first concatenation vector is input into a Bi-LSTM neural network layer to generate a second concatenation vector, the second concatenation vector serving as one of input vectors of a multi-layer graph attention network; the adjacency matrix and the concatenation vector are input into the multi-layer graph attention network as input vectors to obtain an output vector; specifically, the adjacency matrix of the same batch is generated into a tensor; the tensor and the concatenation vector are input into the multi-layer graph attention network for calculation, and the output vector is generated according to the result of the multi-layer graph attention network; aggregation information is generated according to the concatenation vector and the output vector; specifically, in each layer of the graph attention network, the aggregation of the syntactic information is realized through a skip connection module, the concatenation vector is skipped through each layer of the graph attention network through the skip connection module, and an aggregation operation is performed with the output vector; a trigger word category of the context words is determined according to the aggregation information.

2. The event detection method based on the multi-layer graph attention network according to claim 1, wherein, The step of determining the syntactic information adjacency matrix corresponding to the context words comprises: determining the syntactic information corresponding to the context words according to the context words; generating the syntactic information adjacency matrix according to the syntactic information.

3. The method of claim 2, wherein, The step of determining the syntactic information corresponding to the context words according to the context words comprises: analyzing the event text information through syntactic dependency, and generating the syntactic information corresponding to the context words according to the analysis result of the event text information.

4. The method of claim 1, wherein, The step of determining the trigger word category of the context words according to the aggregation information comprises: determining a trigger word of the context words according to the aggregation information, and classifying the trigger word through a classifier module.

5. An event detection apparatus based on a multi-layer graph attention network, characterized in that, The method comprises the steps of: an obtaining module is configured to obtain context words in event text information, and determine a syntactic information adjacency matrix and a concatenation vector corresponding to the context words, including that each sentence in the event text information forms a syntactic dependency tree through syntactic dependency analysis, and then a corresponding adjacency matrix is generated according to the syntactic dependency tree; the concatenation vector is generated according to a word embedding vector of the context words; four different word embedding vectors of a word sense vector, an entity vector, a part-of-speech vector and a position vector of the context words are concatenated into a first concatenation vector, and then the first concatenation vector is input into a Bi-LSTM neural network layer to generate a second concatenation vector, the second concatenation vector serving as one of input vectors of a multi-layer graph attention network; The computing module is configured to take the adjacency matrix and the concatenation vector as inputs of an artificial neural network, and obtain an output vector; specifically, the computing module comprises: an array conversion submodule configured to generate a tensor from the adjacency matrix of the same batch; and an artificial neural network computing submodule configured to input the tensor and the concatenation vector into the artificial neural network for computation, and generate the output vector according to a result of the computation of the artificial neural network; The aggregation module is configured to aggregate the concatenation vector and the output vector to generate aggregated information; specifically, in the graph attention network of each layer, the aggregation of the syntactic information is implemented by a skip connection module, the concatenation vector is made to skip the graph attention network of each layer through the skip connection module, and an aggregation operation is performed on the concatenation vector and the output vector. The classification module is configured to determine a trigger category of the context word according to the aggregated information.

6. The event detection apparatus based on multi-layer graph attention network according to claim 5, characterized in that, The obtaining module comprises: An expression submodule configured to determine syntactic information corresponding to the context word according to the context word; A generation submodule configured to generate the syntactic information adjacency matrix according to the syntactic information.

7. The event detection apparatus based on multi-layer graph attention network according to claim 6, characterized in that, The expression submodule comprises: A dependency analysis submodule configured to analyze the event text information through syntactic dependency, and generate the syntactic information corresponding to the context word according to an analysis result of the event text information. 8.The event detection apparatus based on the multi-layer graph attention network according to claim 5, characterized in that, The classification module comprises: A trigger processing submodule configured to determine a trigger of the context word according to the aggregated information, and classify the trigger according to a classifier module.

Images