Event extraction method and system fusing graph attention network and qa extraction paradigm

Abstract

The application discloses an event extraction method and system fusing a graph attention network and a QA extraction paradigm, and aims at historical combat texts including military events. Based on a graph attention network and a classification mechanism, the historical combat texts are preprocessed, trigger words are recognized, and event argument role extraction is carried out, so that the text is processed, the accuracy of event extraction is improved, and the extraction efficiency of key information is improved and the time of command decision is shortened in a complex and changeable battlefield environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, electronics, communication, information engineering, in particular to an event extraction method fusing a graph attention network and a QA extraction paradigm, and an event extraction system fusing a graph attention network and a QA extraction paradigm. BACKGROUND

[0002] With the "explosive" growth of battlefield information, a large amount of combat data flows into joint combat command institutions, which seriously restricts the timeliness of "OODA" command decision-making, and even causes many key information to be submerged in a large amount of combat data. In the face of multi-domain expansion of combat elements, the battlefield information is deployed in a non-linear, irregular and wide-area manner in a wide space, and an automated and intelligent method is urgently needed to process and utilize combat data and to mine the potential value of combat data. The event extraction technology can automatically extract key events from unstructured combat data and represent them in a structured form, which not only improves the processing and utilization efficiency of combat data and speeds up the combat command decision-making, but also helps the computer to understand the semantics of combat documents to some extent, and has important significance for building combat knowledge graph, mining implicit battlefield information, dynamically updating battlefield situation map and recommending key information for commanders.

[0003] Event extraction is to detect the required event type from the event description text and obtain the event element information and represent the event in a structured form. Generally speaking, event extraction can be divided into two parts, namely event recognition (trigger word recognition) and argument role classification (parameter extraction). Among them, the main goal of event recognition is to identify the corresponding trigger word and judge the event type, and the goal of argument role classification task is to identify the attribute elements corresponding to the event type in the description text according to the event type identified in the previous step. It has important research value in both civilian and military fields.

[0004] At present, the research on event extraction in the military field is relatively less. In the prior art, taking the extraction of military exercise intelligence information as a breakthrough point, a pattern matching based method is used to extract exercise intelligence, and the results show that the pattern matching based method in a specific field can achieve good results, but the development of patterns requires a lot of work, which is difficult to meet the demand of improving the extraction efficiency of key information and shortening the time of command decision-making in complex and variable battlefield environments. SUMMARY

[0005] The present application aims to provide an event extraction method and system fusing a graph attention network and a QA extraction paradigm, which aims to improve the extraction efficiency of key information and shorten the time of "OODA" command decision-making in complex and variable battlefield environments.

[0006] To achieve the above functions, the event extraction method fusing the graph attention network and the QA extraction paradigm is designed, and the following steps S1-S3 are performed on the combat text describing the military event to extract the preset event type in the text and complete the processing of the text:

[0007] Step S1: collect historical combat texts of military events described in natural language, and pre-process the texts;

[0008] Step S2: the text obtained in step S1 is composed of word units, and the text includes event types and trigger word lists corresponding to the event types, respectively, for the text, a trigger word recognition model based on a graph attention mechanism and a binary classification mechanism is constructed, the text is input into the trigger word recognition model, and whether each word unit in the text belongs to the trigger word list corresponding to each event type is determined as the output of the trigger word recognition model, and the identification of the event type and the trigger word in the text is completed;

[0009] Step S3: each event type corresponds to a preset role parameter, for the text whose event type has been determined, a argument role recognition model based on a graph attention mechanism and a Bi-GRU network is constructed, the text and the role parameter corresponding to the event type are input into the argument role recognition model, and the role parameter to which each word unit in the text belongs is output as the output of the argument role recognition model, the identification of the role parameter of the event type is completed, and the processing of the text is further completed.

[0010] The application also designs an event extraction system fusing the graph attention network and the QA extraction paradigm, and based on the event extraction model, the event extraction method fusing the graph attention network and the QA extraction paradigm is realized;

[0011] The event extraction model includes a trigger word recognition model and an argument role recognition model.

[0012] The trigger word recognition model takes the pre-processed text as the input, and takes the determination result of whether each word unit in the text belongs to the trigger word list corresponding to each event type as the output, and includes a coding layer, a context attention layer, and a label embedding layer, a label graph attention layer connected in sequence and in parallel with the above two layers; the trigger word recognition model further includes a label text fusion layer and a binary classification layer; the outputs of the context attention layer and the label graph attention layer are input into the label text fusion layer, the output of the label text fusion layer is taken as the input of the binary classification layer, and the output of the binary classification layer is taken as the output of the trigger word recognition model;

[0013] The argument role recognition model takes the text whose event type has been determined and the role parameter corresponding to the event type as the input, and takes the role parameter to which each word unit in the text belongs as the output, and includes a role entity splicing layer, a coding layer, an attention layer, a Bi-GRU layer, and a classification layer connected in sequence.

[0014] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0015] Compared with existing technologies, this invention proposes an event extraction method and system that integrates graph attention networks and QA extraction paradigms. Given an input sentence, the method first identifies trigger words for the event type by using the event type as the query statement of the sentence. Each event type (e.g., an attack event) is a piece of natural language text, which includes the type name and a list of trigger words corresponding to the type (e.g., attack, assault, etc.). The trigger word list is a set of words representing the event type selected proportionally from the training set. Then, the input sentence and the event type query are fed into a pre-trained BERT model for encoding, and the encoded event type query is interacted with the input text through attention interaction. The number of times events occur simultaneously in the training set is calculated, and the event type is used as a node to connect the simultaneously occurring events. The edge attribute is the number of times several types of simultaneous occurrences occurred. Then, a graph attention network (GAT) is used to fuse information between event types to obtain an event type encoding that integrates information from adjacent events, and then interacts with the input sentence through attention interaction. Finally, multiple encoded vectors are concatenated to perform binary classification on each token to determine whether it is a trigger word for a specific event. For parameter extraction, the same strategy as for trigger word recognition is adopted. Predefined role parameters corresponding to the event type are used as natural language queries and input along with the input sentence into a pre-trained BERT model to learn their contextual representation. A multi-directional attention mechanism is used to obtain the semantic relevance between the input sentence and the role parameters. Then, a Bi-GRU network is used to extract features from the input sentence. Finally, sequence labeling is performed on the input sentence to extract the role parameter entities. Experimental results show that the proposed method effectively improves the accuracy of event extraction. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an event extraction model provided according to an embodiment of the present invention;

[0017] Figure 2 It is the DUEE dataset format provided in the embodiments of the present invention;

[0018] Figure 3 This is a schematic diagram of a graph attention model provided according to an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram of a Bi-GRU network provided according to an embodiment of the present invention;

[0020] Figure 5 This is a comparison chart of the trigger word recognition performance of different methods provided in the embodiments of the present invention;

[0021] Figure 6This is a performance comparison chart of different methods for extracting event role parameters according to embodiments of the present invention;

[0022] Figure 7 This describes the performance of the Roberta pre-trained model provided in the embodiments of the present invention on the trigger word recognition subtask;

[0023] Figure 8 This describes the performance of the Bert-wwm pre-trained model provided in the embodiments of the present invention on the trigger word recognition subtask;

[0024] Figure 9 This is the effect of triggering word sub-tasks when multiple events account for 1 / 2 of the test set, as provided in the embodiments of the present invention;

[0025] Figure 10 This is the effect of triggering word sub-tasks when multiple events account for 1% of the test set, as provided in the embodiments of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0027] The event extraction method fusion graph attention network and QA extraction paradigm provided in this embodiment of the invention is designed for operational texts describing military events, referring to... Figure 1 Perform the following steps S1-S3 to extract the preset event types from the text and complete the text processing:

[0028] Step S1: Collect historical combat texts of military events described in natural language and preprocess the texts;

[0029] Text preprocessing includes word segmentation, data format conversion, and long text truncation.

[0030] Step S2: The text obtained in Step S1 consists of tokens. The text includes various event types and a list of trigger words corresponding to each event type. For the text, a trigger word recognition model based on graph attention and binary classification is constructed. The text is used as the input of the trigger word recognition model, and the judgment result of whether each token in the text belongs to the list of trigger words corresponding to each event type is used as the output of the trigger word recognition model, thus completing the recognition of event types and trigger words in the text.

[0031] The following is an embodiment of the present invention. The event extraction model adopts a pipeline model, which divides event extraction into two parts: trigger word recognition and argument role extraction.

[0032] To validate the constructed model, the DUEE dataset was used to obtain and construct training samples. DUEE is a Chinese event extraction dataset released by Baidu, containing 65 event types and 121 argument roles. The event types and corresponding role parameters of the DUEE dataset are shown in Table 1 below:

[0033] Table 1. Some event types and corresponding role parameters

[0034]

[0035]

[0036] The data format of the DUEE dataset is as follows: Figure 2 Here, “text” represents the text content, “event_list” represents the events contained in the text, “event_type” represents the event type, “trigger” represents the trigger word, “trigger_start_index” represents the trigger word position, and “argument” represents the role parameters, including role type, position, and role entity.

[0037] The trigger word recognition model works as follows in step S2:

[0038] Step S21: Given an input sentence, identify trigger words for the event type by using the event type as the query statement of the sentence. Each event type (e.g., attack event) is a piece of natural language text that includes the type name and a list of trigger words corresponding to the type (e.g., attack, assault, etc.). The trigger word list is a set of words representing the event type selected proportionally from the training set. Input the text and the query for each event type into the pre-trained BERT model for encoding;

[0039] This approach differs from traditional methods that directly input the raw text into the BERT model for training, ignoring the semantic information carried by event type labels. For example, in the sentence "Chinese athletes defeated their opponents in the marathon and successfully advanced to the semifinals," both "advanced" and "defeated" could refer to the Chinese athletes winning the marathon. The existence of the "advanced" event increases the probability of the "defeated" event, and vice versa. Considering the mutual influence between events, it is necessary to construct a graph of event types and use a graph attention mechanism to fuse the relationships between events. To further enhance the semantic meaning of the labels, event types and common trigger words are connected as prior knowledge. For example, common trigger words for the "advanced" event are "advanced," "qualified," "advanced," and "breakthrough," resulting in the format [CLS]-advanced-[SEP]-advanced-qualified-advanced-breakthrough-[SEP]". This format is then input into the BERT model along with the raw text for training.

[0040] Step S22: Given a query for each event type, the goal is to automatically extract the corresponding event trigger words from the input sentence. To achieve this, it is necessary to capture the semantic association between each input token and the event type.

[0041] For each encoded event type query and text, a graph attention mechanism is used to learn the weight distribution on the context representation sequence of the event type query, and to obtain the context representation of each word in the text after fusing the event type, as shown in the following formula:

[0042]

[0043]

[0044] In the formula, Let be the vector representation of the i-th word in the text. Let α be the vector representation of the j-th event type in the text, cos(·) be the cosine similarity value between the two vectors, and α be the vector similarity value between the two vectors. ij Let be the cosine similarity matrix between the i-th word and the j-th event type; Let k be the context representation of the event type after fusing the i-th lexical element, where k is the total number of event types.

[0045] Furthermore, the identification of event trigger words is also influenced by specific contexts. For example, the context of an attack event may include words such as names, locations, and times. To capture such contextual information, a contextual attention mechanism is further employed to capture each piece of contextual information in the text, as shown in the following formula:

[0046]

[0047]

[0048] In the formula, ρ(·) is the attention function, and β ij Let be the similarity coefficient between the i-th and j-th word elements in the text, and N be the total number of word elements in the text. A lexical vector representation that incorporates information from adjacent lexical units;

[0049] Step S23: For each text, using the event types contained within it as label nodes, connect the event types that appear simultaneously with each other, with the edge attribute representing the number of times each event type appears simultaneously. Since each text contains multiple label nodes, there may be dependencies or correlations between label nodes in different texts. Based on this, according to the co-occurrence principle, construct a graph G = (V, E) to mine the relationships between labels, where V represents label nodes, E represents edges, and the adjacency matrix A is formed by the topological structure of graph G. ij As shown in the following formula:

[0050]

[0051]

[0052]

[0053] In the formula, n ij C is the number of times two label nodes i and j co-occur in all text. m This represents the m-th text. This indicates that matrix X ij All diagonal elements are set to 1, representing a self-loop operation for each tag node. I is the identity matrix, and D is the matrix... The degree matrix, A ij Soon Matrix Perform normalization processing;

[0054] Step S24: Use as follows Figure 3 The graph attention network (GAT) shown fuses information between event types to obtain event type codes that incorporate information from adjacent events, and then interacts with the input sentence through attention. Its graph attention mechanism is as follows:

[0055] Based on the adjacency matrix A ij And using GloVe pre-trained word vectors to represent the initial features of the label nodes. k is the number of label nodes, d is the number of features for each label node, and a graph attention mechanism is used to aggregate each label node with its first-order neighbor label nodes. Dynamic attention is then used to obtain the feature representation of each label node. This also further strengthens the correlation between tag nodes, as shown in the following formula:

[0056]

[0057]

[0058] In the formula, Let W be the weight parameter vector, W be the weight parameter matrix, LeakyReLU be the non-linear activation function, and α be the weight parameter vector. ij Let j be the normalized attention coefficient of label node j relative to label node i, k∈N i Let f represent all first-order neighboring label nodes of label node i, and f be a non-linear activation function.

[0059] The label information obtained by the graph attention network in steps S22 and S24 is concatenated to obtain the label features.

[0060]

[0061] In the formula, t represents the number of layers in the graph attention network (GAT), and || is the concatenation symbol. W represents the normalized attention coefficient of label node j relative to label node i in the t-th operation of the graph attention network GAT. t Let f be the weight parameter matrix in the t-th operation of the graph attention network GAT, and f be the non-linear activation function.

[0062] Step S25: Concatenate multiple encoded vectors and perform binary classification on each token to determine whether it is a trigger word for a specific event type. The classification process is as follows:

[0063] The results obtained from steps S22 and S24 through a graph attention network It is concatenated with the original text context representation obtained through a pre-trained BERT model and classified into a binary label indicating whether it is a trigger word corresponding to event type t:

[0064]

[0065] In the formula, U0 is the learnable parameter matrix for event-triggered detection, and P is the parameter matrix for event-triggered detection. i It is the word W i Single-hot part-of-speech encoding;

[0066] Its loss function is:

[0067]

[0068] In the formula, |T| represents the number of event types, and |N| represents the length of the input text. For correct labeling, The predicted label.

[0069] Step S3: Each event type corresponds to a preset role parameter. For text with a determined event type, an argument role recognition model based on graph attention mechanism and Bi-GRU network is constructed. The role parameters corresponding to the text and event type are used as the input of the argument role recognition model, and the role parameters of the words in the text are used as the output of the argument role recognition model. The role parameters of the event type are identified, and the text is further processed.

[0070] Using the same strategy as the trigger word recognition model, argument roles are used as labels. Predefined role parameters corresponding to the event type are input along with the natural language query and the input sentence into the pre-trained BERT model to learn its contextual representation. The working process of the argument role recognition model in step S3 is as follows:

[0071] Step S31: Given input text C = [c1, c2, c3, ..., c] with a defined event type. n ], set the preset role parameter A = [a1, a2, ..., a] belonging to this event type. k The vector is formed by concatenating the input text with the vector shown below:

[0072] [CLS]a1,a2,...,a k [SEP]c1,c2,c3,…,c n [SEP]

[0073] The vectors are input into the BERT pre-trained model to encode each sentence in the text, resulting in a vector representation that incorporates contextual semantics, as follows:

[0074]

[0075] In the formula, W represents the matrix composed of word vectors. Represents the word vectors of event role parameters. This represents the text word vector; k+n represents the length of the concatenation of the text and the character parameters.

[0076] Step S32: Use a multi-directional attention mechanism to obtain the semantic relevance between each sentence in the input text and the role parameters, where the text-role parameter attention is as follows:

[0077] Given a text vector w c With the character parameter vector w a To determine the correlation between text and character parameters, a similarity matrix S between text and character parameters is calculated. ij :

[0078]

[0079] In the formula, d represents w c with w a The vector dimensions are the same for both vectors.

[0080] Based on the similarity matrix S ij The attention mechanism is used to obtain text vectors that incorporate semantic information of character parameters. Role parameter vectors with fused text information

[0081]

[0082]

[0083] To extract latent relationships between text lexical units or parameter roles, self-attention is calculated for both the text and the parameter roles:

[0084]

[0085] η = Softmax(η) ij )

[0086]

[0087]

[0088] μ = Softmax(μ ij )

[0089]

[0090] In the formula, σ represents the dot product operation; η ij This represents the semantic relevance between the i-th and j-th words in the text; Softmax(·) represents the normalization function; η represents ηj. ij Normalized representation; μ represents the vector representation of the i-th lexical unit in the text that incorporates contextual information. ij μ represents the semantic relevance between the i-th argument and the j-th argument; μ represents μ ij Normalized representation; The i-th argument role vector representation that incorporates information from other argument roles;

[0091] Step S33: Extract features of the input sentence using a Bi-GRU network, the process of which is as follows:

[0092] Recurrent Neural Networks (RNNs) capture the relationship between the current state and previous states through a network structure of repetitive execution units, enabling them to process time series of arbitrary length. GRU networks are simplified variants of Long Short-Term Memory (LSTM) networks, reducing computational overhead by eliminating one hidden state while achieving performance similar to LSTM. To improve the real-time performance of perception, this invention uses a bidirectional GRU network to extract global features of text and role parameters, as illustrated in the diagram below. Figure 4 In a GRU network, there is a hidden state h t The two gate control units are for resetting the gate r t and update gate z t and candidate status The specific calculation method is as follows:

[0093] r t =sigmoid(W r x t +U r h t-1 )

[0094] z t =sigmoid(W z x t +U z h t-1 )

[0095]

[0096]

[0097] In the GRU gated unit, short-term gate control dependencies are reset, and long-term gate control dependencies are updated. The input is the feature x at the current time step. t and the hidden state h from the previous moment t-1 The output activation function is the sigmoid function, which maps the output to a range of 0 to 1, controlling the inflow of information. Candidate states. The dependency between the current input and previous states is controlled by a reset gate. If the reset gate approaches 0, the previous states are completely forgotten, and the candidate states are equivalent to the current input information. The current state h... t This is also the network output, obtained by linear interpolation of the candidate state and the previous time step state. The bidirectional GRU network includes forward hidden states and backward hidden states, and the calculation method for each direction is the same as described above. The bidirectional network can capture the forward and backward connections of the spectral data better.

[0098] The global features captured by the GRU network are reduced by multiplicative attention to mitigate the impact of noise, calculated as follows:

[0099] S = tanh(h) t )

[0100] α = softmax(w T S)

[0101] O = h t ·α T

[0102] Where S represents the attentional hidden state; α represents the attention coefficient; w T represents the learnable parameter matrix; O represents the text vector representation that incorporates global features;

[0103] Step S34: Input the output value of the GRU network into the fully connected layer for sequence labeling to obtain the role parameter entity, as shown in the following formula:

[0104]

[0105] In the formula, U0 is the learnable parameter matrix for event-triggered detection;

[0106] Its loss function is:

[0107]

[0108] In the formula, |A| represents the role parameter of the event type. For the correct parameter role label, The predicted parameter is the role label.

[0109] This invention also provides an event extraction system that integrates graph attention networks and QA extraction paradigms, and implements the event extraction method that integrates graph attention networks and QA extraction paradigms based on the event extraction model.

[0110] Reference Figure 1 The event extraction model includes a trigger word recognition model and an argument role recognition model;

[0111] The trigger word recognition model takes preprocessed text as input and outputs the result of whether each word in the text belongs to the trigger word list corresponding to each event type. It includes an encoding layer and a context attention layer connected in sequence, as well as a label embedding layer and a label graph attention layer connected in parallel and in sequence with the above two layers. The trigger word recognition model also includes a label text fusion layer and a binary classification layer. The outputs of the context attention layer and the label graph attention layer are respectively input to the label text fusion layer, the output of the label text fusion layer is used as the input of the binary classification layer, and the output of the binary classification layer is used as the output of the trigger word recognition model.

[0112] The argument role recognition model takes text with a defined event type and the corresponding role parameters as input, and outputs the role parameters to which the words in the text belong. It includes a role entity concatenation layer, an encoding layer, an attention layer, a Bi-GRU layer, and a classification layer connected in sequence.

[0113] The role entity concatenation layer of the argument role recognition model is used to concatenate the pre-defined role parameters belonging to that event type with the input text to form a vector for a given input text of a defined event type. The encoding layer inputs the vector into the BERT pre-trained model to encode each sentence in the text, obtaining a vector that incorporates contextual semantics. The attention layer interacts with the contextual semantic vector through text-role parameter attention and text-text and role-role attention to obtain a text vector representation that incorporates multi-directional attention. The Bi-GRU layer takes the output of the attention layer as the input of the Bi-GRU to obtain a text vector representation that incorporates long sequence information. The classification layer feeds the output text vector into the fully connected layer for multi-classification to obtain the entity of the event role parameters.

[0114] Figure 5 The performance comparison of trigger word recognition using different methods is described. Specific data is shown in Table 2:

[0115] Table 2 Trigger Word Recognition

[0116]

[0117] Comparative experiments show that, compared to the BERT_QA_Arg model, the QA-GAT model improves the accuracy of event extraction by 6.09%, the recall by 2.72%, and the F1 score by 4.49% in trigger word recognition.

[0118] Figure 6 The performance comparison of event role parameter extraction using different methods is described. Specific data is shown in Table 3.

[0119] Table 3 Argument Role Recognition

[0120]

[0121] Comparative experiments show that, compared with the BERT_QA_Arg model, the QA-GAT model improves the accuracy of event extraction by 10.9%, the recall by 9.11%, and the F1 score by 10.02% in argument role recognition.

[0122] Figure 7 and Figure 8 The graphs show the F1 score and loss variation of the Roberta model and the Bert-WWM model on the trigger word recognition subtask, respectively. Figure 7 and Figure 8 The comparison shows that the Bert-wwm model is better than the Roberta model in both convergence speed and F1 score. Therefore, the Bert-wwm model is chosen for word embedding.

[0123] Figure 9 and Figure 10 These are the F1 scores on the trigger word recognition subtask when the proportion of multiple events in the test set is 1 / 2 and 1, respectively. Figure 9 and Figure 10 The comparison shows that the method proposed in this invention has an effective improvement in multi-event extraction and can extract multiple events better.

[0124] In summary, this paper analyzes the advantages and disadvantages of several existing event extraction methods from the perspective of operational text event extraction. Based on this analysis, an event extraction model integrating graph attention networks and QA extraction paradigms is proposed, which improves operational efficiency and significantly increases accuracy to a certain extent. Therefore, the event extraction model proposed in this invention is feasible and effective.

[0125] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An event extraction method fusing graph attention networks and QA extraction paradigm, characterized in that, For the combat text describing the military event, the following steps S1-S3 are performed to extract the preset event type in the text and complete the processing of the text: Step S1: collect historical combat text of a military event described in natural language, and pre-process the text; Step S2: the text obtained in step S1 is composed of word units, and the text includes event types and trigger word lists corresponding to the event types respectively, for the text, a trigger word recognition model based on a graph attention mechanism and a binary classification mechanism is constructed, the text is input into the trigger word recognition model, and whether each word unit in the text belongs to the trigger word list corresponding to each event type is determined as the output of the trigger word recognition model, so that the identification of the event type and the trigger word in the text is completed; Step S3: each event type corresponds to a preset role parameter, for the text whose event type has been determined, a argument role recognition model based on a graph attention mechanism and a Bi-GRU network is constructed, the text and the role parameter corresponding to the event type are input into the argument role recognition model, and the role parameter to which each word unit in the text belongs is output from the argument role recognition model, so that the identification of the role parameter of the event type is completed, and the processing of the text is further completed; The working process of the argument role recognition model is as follows: Step S31: Given the input text of a certain event type , the preset role parameter of the event type is concatenated with the input text to form the following vector: ； The vector is input into the BERT pre-training model to encode each sentence in the text, and a vector representation fusing the context semantics is obtained as follows: ； In the formula, W represents a matrix composed of word vectors, represents an event role parameter word vector, represents a text word vector; k+n represents the length of the concatenation of the text and the role parameter; Step S32: using a multi-directional attention mechanism, the semantic correlation between each sentence in the input text and the role parameter is obtained; Step S33: using a Bi-GRU network to extract the features of the input sentence; Step S34: the output value of the GRU network is put into a fully connected layer for sequence labeling, and the role parameter entity is obtained.

2. The method of claim 1, wherein the fusion graph attention network and QA extraction paradigm for event extraction is characterized by, The working process of the trigger word recognition model in step S2 is as follows: Step S21: the text and each event type query are input into a pre-trained BERT model for encoding; Step S22: for the encoded event type query and text, a graph attention mechanism is used to learn the weight distribution on the context representation sequence of the event type query, and the context representation of each word unit in the text after fusing the event type is obtained, which is specifically as follows: ； ； wherein, is a vector representation of the i-th wordpiece in the text, is a vector representation of the j-th event type in the text, is a cosine similarity value of two vectors, is a cosine similarity matrix of the i-th wordpiece and the j-th event type; is a contextual representation of the i-th wordpiece after merging event types, k is the total number of event types. Further, the context attention mechanism is used to capture each context information in the text, which is specifically as follows: ； ； wherein is an attention function, is a similarity coefficient between the i-th token and the j-th token in the text, and N is the total number of tokens included in the text, denotes a token vector representation that fuses the information of adjacent tokens; Step S23: For the text, taking each event type contained in the text as a label node, and connecting edges between simultaneously occurring event types, a graph is constructed wherein V represents a label node, E represents an edge, and an adjacency matrix composed of the topological structure of the graph G as follows: ； ； ； wherein, is the number of times two label nodes i, j appear together in all texts, denotes the mth text, denotes setting all diagonal elements of the matrix to 1, I is the identity matrix, and D is the degree matrix of the matrix denotes the normalized adjacency matrix;​ Step S24: according to the adjacency matrix and the initialization features of the label nodes are represented using the Glove pre-trained word vectors , k is the number of label nodes, d is the number of features of each label node, the graph attention mechanism is used to aggregate each label node and its first-order neighborhood label nodes, and dynamic attention is used to obtain the feature representation of each label node , as follows: ； ； wherein, is a vector of weight parameters, W is a matrix of weight parameters, is a nonlinear activation function, is a normalized attention coefficient of label node j with respect to label node i, denotes all first-order neighborhood label nodes of label node i, is a nonlinear activation function; The label information representation obtained by the graph attention network of step S22 and step S24 is spliced to obtain label features : ； In the formula, t represents the number of layers of the graph attention network GAT, and || is a splicing symbol, is a normalized attention coefficient of the label node j with respect to the label node i in the t-th operation of the graph attention network GAT, is a weight parameter matrix of the graph attention network GAT in the t-th operation, is a nonlinear activation function; Step S25: the text context representation obtained by the graph attention network in step S22 and step S24 is connected with the original text context representation obtained by the pre-trained BERT model, and is classified into a binary label indicating whether it is a trigger word corresponding to the event type t: , , and the original text context representation obtained by the pre-trained BERT model, and is classified into a binary label indicating whether it is a trigger word corresponding to the event type t: ； wherein learnable parameter matrix of event-triggered detection, is a one-hot word encoding of the word ​ The loss function is: ； wherein, is the number of event types, is the length of the input text, is the correct label, is the predicted label.

3. The fusion graph attention network and QA extraction paradigm based event extraction method according to claim 2, characterized in that, The working processes of steps S32-S34 in step S3 are as follows: The process of step S32 is as follows: Given text vector With character parameter vector To determine the correlation between the text and the character parameters, a similarity matrix between the text and the character parameters is calculated : ； In the formula, d represents with vector dimension; Based on similarity matrix , the attention mechanism is used to obtain the text vector of the fusion role parameter semantic information and the role parameter vector of the fusion text information : ； ； Self-attention calculation is performed on the text and the parameter role: ； ； ； ； ； ； wherein denotes a dot product operation; denotes a semantic correlation between the ith wordpiece and the jth wordpiece in the text; denotes a normalization function; denotes a normalized representation of denotes an ith wordpiece vector representation incorporating context information in the text; denotes a semantic correlation between the ith argument role and the jth argument role; denotes a normalized representation of an ith argument role vector representation incorporating other argument role information; The process of step S33 is as follows: There is a hidden state in the GRU network Two gating units are reset gate and update gate and candidate state The specific calculation is as follows: ； ； ； ； In the formula, is the hidden state of the previous time, is the feature of the current time, represents the control reset gate, represents the control update gate, represents the trainable weight matrix of the current time feature in the reset gate, represents the trainable weight matrix of the previous time hidden state in the reset gate, represents the trainable weight matrix of the current time feature in the update gate, represents the trainable weight matrix of the previous time hidden state in the update gate, represents the splicing of the previous state after the reset gate is reset and the current time feature, used to remember the current time state, represents the trainable weight matrix of the current time feature, represents the trainable weight matrix of the previous time hidden state after the reset gate is reset. The global features captured by the GRU network are reduced by multiplicative attention to reduce the noise influence, and the calculation method is as follows: ； ； ； wherein, is an attention hidden state; is an attention coefficient; denotes a learnable parameter matrix; O denotes a text vector representation that fuses global features; The process of step S34 is as follows: ； In the formula, is a learnable parameter matrix for event-triggered detection; The loss function is: ； where |A| is the argument of the event type, is the correct argument role label, is the predicted argument role label.

4. An event extraction system that fuses graph attention networks and the QA extraction paradigm, characterized in that, Based on the event extraction model, the event extraction method of fusing the graph attention network and the QA extraction paradigm is realized according to any one of claims 1-3; The event extraction model includes a trigger word recognition model and an argument role recognition model; The trigger word recognition model takes the preprocessed text as input and takes the judgment result of whether each word element in the text belongs to the trigger word list corresponding to each event type as output, including sequentially connected encoding layers, context attention layers, and label embedding layers, label graph attention layers connected in parallel with the above two layers; the trigger word recognition model further includes a label text fusion layer and a binary classification layer; the outputs of the context attention layer and the label graph attention layer are respectively input into the label text fusion layer, the output of the label text fusion layer is taken as the input of the binary classification layer, and the output of the binary classification layer is taken as the output of the trigger word recognition model; The argument role recognition model takes the text whose event type has been determined and the role parameters corresponding to the event type as input, and takes the role parameters to which the word elements in the text belong as output, including sequentially connected role entity splicing layers, encoding layers, attention layers, Bi-GRU layers, and classification layers.

Images