An event extraction method, event extraction framework and system based on semantic dependency analysis

Abstract

The application provides an event extraction method based on semantic dependency analysis, and relates to the technical field of natural language processing. A simplified text graph is constructed through semantic dependency analysis, the simplified text graph is divided based on a meta path to obtain a meta path subgraph, in an event argument extraction process, a cross-view comparison learning is performed on the simplified text graph and the meta path subgraph respectively by using a graph convolution network to extract event argument features, meanwhile, in an event relation extraction process, an alternating learning is performed on each adjacent event relation by using an event relation learner, finally, based on the trained event argument extraction process and the event relation extraction process, event arguments and event relations are jointly extracted from to-be-identified text. The application extracts event argument features based on semantic dependency relations, and simultaneously implements extraction of event relations, realizes joint extraction of event arguments and event relations, and improves the accuracy of event extraction.

Description

Technical Field

[0001] This invention belongs to the technical field of natural language processing, and more specifically relates to an event extraction method, event extraction framework and system based on semantic dependency analysis. Background Technology

[0002] Event extraction is an important task in natural language processing (NLP), a crucial method for extracting basic information from text. It transforms unstructured data in text into structured data, concisely and clearly representing the structure of events. Event extraction can be divided into two parts: event argument extraction and event relation extraction. Event arguments identify entities related to specific events from the text, while event relations identify semantic connections between events (such as temporal, causal, and coreference). By extracting event arguments and event relations from the text and converting them into event triples in the form of argument-relation-argument, it enables the analysis and identification of event elements such as news events, financial events, and medical events. It can also be used to construct information systems and visualize knowledge graphs, facilitating information retrieval and management.

[0003] In recent years, deep learning technology has developed rapidly, and its application in event extraction has also seen remarkable progress. Currently, various deep network models for event extraction have been developed, such as DMCNN (Dynamic Multi-Pooling Convolutional Neural Network), JRNN (Joint Recurrent Neural Network), DBRNN (Dependency-Bridge RNN), and Transformers (a general deep learning architecture based on self-attention). At present, event extraction methods mainly include pipeline-based methods and joint learning-based methods. Pipeline-based methods transform the event extraction task into multi-stage sub-tasks, first identifying the trigger type, and then extracting parameters such as event arguments and event relationships based on the triggers. However, this staged strategy can lead to error propagation. Joint learning-based methods extract triggers and other event parameters simultaneously, effectively preventing error propagation; however, their performance degrades in complex event structures.

[0004] Semantic dependency is a fundamental relationship between words, effectively representing the syntactic structure of text. Event arguments typically contain rich dependencies and exhibit similarities, all of which can be helpful for event extraction tasks. However, most existing event extraction methods model word relationship networks based on the contextual semantic information of the text, neglecting the relevant information between words. While textual contextual information can represent some relationships between words, these relationships are superficial. The most direct manifestation of dependencies between event elements is semantic dependency, which expresses relevant information between words. Arguments within the same event typically have very strong semantic dependencies. Analyzing semantic dependencies helps to gain a deeper understanding of the structured information of the text, thus more clearly identifying the potential event structure information within the text. However, in practical applications, due to the complexity of events themselves and the complex semantic dependency structure between words within events, event arguments and event relationships may need to be extracted sequentially, leading to error propagation. If the event argument extraction result is incorrect, the event relationship may be further misinterpreted, affecting the accuracy of event extraction. Summary of the Invention

[0005] To address the low accuracy of existing event extraction methods based on semantic dependency analysis, this invention provides an event extraction method, framework, and system based on semantic dependency analysis, which jointly extracts event arguments and event relations, thereby improving the accuracy of event extraction.

[0006] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows:

[0007] S1: Encode the text to obtain the text embedding vector;

[0008] S2: Perform semantic dependency analysis on the text, and construct a text semantic dependency graph structure based on the semantic dependency analysis results. The text semantic dependency graph structure includes nodes and semantic dependency edges. Simplify the semantic dependency edges to obtain a simplified text graph.

[0009] S3: Divide the simplified text graph based on the meta-path to obtain the meta-path sub-graph;

[0010] S4: Use graph convolutional networks to perform cross-view comparative learning on simplified text graphs and metapath subgraphs respectively, and obtain node features under simplified text graph view and node features under metapath view respectively;

[0011] S5: Map the node features under the simplified text graph view and the node features under the meta-path view to the same feature space, calculate the contrast loss, and use the contrast loss as the loss function for training the event argument extraction process.

[0012] S6: Input the text embedding vector into the event relation learner, and use the event relation learner to learn each adjacent event relation alternately to train the event relation extraction process;

[0013] S7: Based on the event argument extraction process trained in S5 and the event relation extraction process trained in S6, jointly extract event arguments and event relations from the text to be identified.

[0014] Furthermore, the text is encoded using the BERT model to obtain the text embedding vector, where the text is represented as S = {w1, w2, ..., w...} n}, w n Let H represent the nth word in the text. The embedding vector of the text is: H = {v1, v2, ..., v...} n}, v n This represents the embedding vector of the nth word.

[0015] Furthermore, semantic dependency analysis tools are used to perform semantic dependency analysis on the text, and the semantic dependency analysis results are obtained. The text semantic dependency graph structure is constructed as R=(W,Rela), where W={m1,m2,…,m n} represents a set of nodes, and Rela represents a set of semantic dependencies;

[0016] Different types of edges are defined to represent the relationships between different words. Each type of edge has a unique identifier, and the expression is:

[0017]

[0018] In the formula, ID edge The edge number is indicated by: Trigger-Triggeredge indicates that both ends of the edge are trigger words; Trigger-Entityedge indicates that both ends of the edge are trigger words and words connected to the trigger words; Entity-Entity-e edge indicates that both ends of the edge are event arguments; and Entity-Entity-n edge indicates that both ends of the edge are event arguments and non-event arguments.

[0019] Construct a grid T of size n×n. For each element T in T... i,j That is, the relationship between the i-th word and the j-th word, which is filled using the edge numbers of the corresponding types, expressed as:

[0020]

[0021] In the formula, T i,j This indicates the relationship between the i-th word and the j-th word. A label indicating the relationship between the i-th word and the j-th word;

[0022] After filling, a simplified text graph G = (W, E) is obtained, where W represents the set of nodes in the simplified text graph and E represents the set of edges in the simplified text graph, i.e., the elements of the grid T.

[0023] Based on the aforementioned technical methods, semantic dependency analysis is used to capture the semantic relationships between words in the text. The simplified text graph can clearly express the core semantic structure of the text, which helps to improve the accuracy of event argument extraction.

[0024] Furthermore, the meta-path subgraph includes: ETE meta-path subgraph, EOE meta-path subgraph, and ETO meta-path subgraph;

[0025] The simplified text graph is divided according to the ETE path to obtain the ETE meta-path subgraph SG. ETE The ETE path consists of nodes that are, in sequence, event arguments, trigger words, and event arguments.

[0026] The simplified text graph is divided according to the EOE path to obtain the EOE metapath subgraph SG. EOE The EOE path consists of nodes that are, in order, event arguments, non-event arguments, and event arguments.

[0027] The simplified text graph is divided according to the ETO path to obtain the ETO metapath subgraph SG. ETO The ETO path consists of nodes that are, in order, event arguments, trigger words, and non-event arguments.

[0028] Finally, we obtain the subgraph set SG = {SG} based on metapaths. ETE ,SG EOE ,SG ETO}

[0029] Furthermore, the process of using graph convolutional networks to perform cross-view comparative learning on the simplified text graph and the meta-path subgraph to obtain node features under the simplified text graph view and the meta-path view respectively is as follows:

[0030] First, construct the adjacency matrix of the simplified text graph; then, select the nodes of the metapath subgraph and their corresponding edges in the adjacency matrix, and use the selected nodes and their corresponding edges as the adjacency matrix of the metapath subgraph.

[0031] Next, the adjacency matrix of the simplified text graph, the adjacency matrix of the metapath subgraph, and the text embedding vector are input into the graph convolutional network to obtain the node features under the simplified text graph view and the node features under the metapath view, expressed as follows:

[0032] L G =ρ(A G HW0)

[0033] L M =ρ(A M HW0)

[0034] In the formula, L G L represents the node features in a simplified text graph view. M Represents the node features under the meta-path view, ρ(·) represents the activation function, W0 represents the weight matrix, and A G Let A represent the adjacency matrix of the simplified text graph. M H represents the adjacency matrix of the metapath subgraph, and H represents the text embedding vector.

[0035] Furthermore, by utilizing MLP, node features in the simplified text graph view and node features in the meta-path view are mapped to the same feature space, resulting in... and The expressions are as follows:

[0036]

[0037] In the formula, This represents the text graph node features after MLP transformation. W represents the node features of the metapath subgraph after MLP transformation. 1 Let W represent the weight matrix. 2 Let b represent the weight matrix. 1 b represents the bias term. 2 Indicates the bias term;

[0038] The node features in the meta-path view mapped to the same feature space are used as target features, and the node features in the simplified text graph view are used as contrast features. The contrast loss is calculated and used as the loss function of the graph convolutional network.

[0039] The comparison loss is divided into three parts: ETE path loss, EOE path loss and ETO path loss.

[0040] The ETE path loss is calculated by taking the node features of the ETE meta-path subgraph as positive samples and the node features of the remaining meta-path subgraphs as negative samples. The expression is as follows:

[0041]

[0042] In the formula, SG represents the ETE path loss. ETO Represents the ETO metapath subgraph, SG ETE Metapath Subgraph (ETE), SG EoEdenoted as EOE metapath subgraph, exp represents the exponential function, and sim represents similarity; Indicates positive sample pairs in the output. and similarity, express The sum of similarities with all negative samples;

[0043] The EOE path loss is calculated by taking the node features of the ETE meta-path subgraph as positive samples and the node features of the remaining meta-path subgraphs as negative samples. The expression is as follows:

[0044]

[0045] In the formula, Indicates EOE path loss;

[0046] The ETO path loss is calculated by taking the node features of the ETO meta-path subgraph as positive samples and the node features of the remaining meta-path subgraphs as negative samples. The expression is as follows:

[0047]

[0048] The expression for the contrast loss is:

[0049]

[0050] In the formula, Indicates comparative loss, Indicates ETE path loss, This represents the EOE path loss. This represents the ETO path loss.

[0051] Based on the aforementioned technical methods, simplified text graphs and meta-path subgraphs capture semantic information of text from different perspectives. Through cross-view comparative learning, the features of these two views can be utilized to improve the ability to capture semantic information. By minimizing the contrastive loss, the event argument extraction process is trained to improve the accuracy of event argument extraction.

[0052] Furthermore, the text embedding vector is input into an event relation learner, which is constructed using an MLP. The event relation learner learns each adjacent event relation alternately; the event relation learner learns the i-th event relation to obtain the output. The expression is:

[0053]

[0054] In the formula, σ(·) represents the activation function, W represents the weight matrix, and b represents the bias term. H represents the predicted result of the event relationship output, where H represents the text embedding vector.

[0055] Furthermore, during the training of the event relationship extraction process, a loss function for calculating the event relationship learner is constructed, and the parameters of the event relationship learner are optimized based on the loss function;

[0056] The loss function of the event relation learner is calculated as follows:

[0057]

[0058] In the formula, This represents the loss function of the event-relationship learner. y represents the predicted result of the event relationship in the output. i Indicates the true outcome of the relationship between events;

[0059] The expression for the parameter optimization process is:

[0060]

[0061] In the formula, θ t θ represents the parameters of the optimized event relation learner. t-1 This represents the parameters of the current event relation learner, where γ represents the learning rate. This represents the gradient of the loss function with respect to the parameters.

[0062] The present invention also provides an event extraction framework based on semantic dependency analysis, which is used to implement the event argument extraction process and the event relation extraction process, including: an event argument extraction subframe and an event relation extraction subframe;

[0063] The event argument extraction sub-framework includes: semantic dependency analysis structure, text graph construction structure, meta-path subgraph structure, graph convolutional network structure, and MLP structure;

[0064] The event relationship extraction sub-framework includes a text encoder structure and an event relationship learner structure.

[0065] This invention also provides an event extraction system based on semantic dependency analysis, comprising:

[0066] The text encoding module is used to encode text and obtain the text embedding vector;

[0067] A simplified text graph construction module is used to perform semantic dependency analysis on text. Based on the semantic dependency analysis results, a text semantic dependency graph structure is constructed. The text semantic dependency graph structure includes nodes and semantic dependency edges. The semantic dependency edges are simplified to obtain a simplified text graph.

[0068] The metapath subgraph construction module is used to divide the simplified text graph based on metapaths to obtain metapath subgraphs;

[0069] The node feature output module is used to perform cross-view comparative learning on the simplified text graph and the meta-path subgraph using a graph convolutional network, respectively, to obtain node features under the simplified text graph view and node features under the meta-path view.

[0070] The event argument extraction process training module is used to map the node features under the simplified text graph view and the node features under the meta-path view to the same feature space, calculate the contrastive loss, and use the contrastive loss as the loss function for training the event argument extraction process.

[0071] The event relationship extraction and training module is used to input the text embedding vector into the event relationship learner, and use the event relationship learner to learn each adjacent event relationship alternately to train the event relationship extraction process;

[0072] The joint extraction module is used to jointly extract event arguments and event relationships from the text to be identified based on the trained event argument extraction process and the trained event relationship extraction process.

[0073] Compared with existing technologies, the beneficial effects of this method are:

[0074] This invention proposes an event extraction method, framework, and system based on semantic dependency analysis. A simplified text graph is constructed through semantic dependency analysis. This graph is then divided into meta-path subgraphs based on meta-paths. During event argument extraction, a graph convolutional network is used to perform cross-view comparative learning on both the simplified text graph and the meta-path subgraphs to extract event argument features. Simultaneously, during event relation extraction, an event relation learner is used to alternately learn each adjacent event relation. Finally, based on the trained event argument extraction and event relation extraction processes, event arguments and event relations are jointly extracted from the text to be identified. This invention extracts event argument features based on semantic dependency relationships and simultaneously extracts event relations, achieving joint extraction of event arguments and event relations and improving the accuracy of event extraction. Attached Figure Description

[0075] Figure 1 A flowchart illustrating the event extraction method based on semantic dependency analysis proposed in this embodiment of the invention;

[0076] Figure 2 This diagram illustrates a detailed flowchart of the event extraction method based on semantic dependency analysis proposed in this embodiment of the invention.

[0077] Figure 3A flowchart illustrating the event argument extraction process proposed in this embodiment of the invention;

[0078] Figure 4 A flowchart illustrating the event relationship extraction process proposed in this embodiment of the invention;

[0079] Figure 5 This is a framework diagram illustrating the event extraction framework based on semantic dependency analysis proposed in this embodiment of the invention.

[0080] Figure 6 This diagram illustrates the structure of the event extraction system based on semantic dependency analysis proposed in this embodiment of the invention. Detailed Implementation

[0081] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0082] To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions;

[0083] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0084] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0085] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0086] Example 1

[0087] This embodiment proposes an event extraction method based on semantic dependency analysis, such as... Figure 1 The flowchart shown illustrates the method, and the method proposed in this embodiment generally includes the following steps:

[0088] S1: Encode the text to obtain the text embedding vector;

[0089] S2: Perform semantic dependency analysis on the text, and construct a text semantic dependency graph structure based on the semantic dependency analysis results. The text semantic dependency graph structure includes nodes and semantic dependency edges. Simplify the semantic dependency edges to obtain a simplified text graph.

[0090] S3: Divide the simplified text graph based on the meta-path to obtain the meta-path sub-graph;

[0091] S4: Use graph convolutional networks to perform cross-view comparative learning on simplified text graphs and metapath subgraphs respectively, and obtain node features under simplified text graph view and node features under metapath view respectively;

[0092] S5: Map the node features under the simplified text graph view and the node features under the meta-path view to the same feature space, calculate the contrast loss, and use the contrast loss as the loss function for training the event argument extraction process.

[0093] S6: Input the text embedding vector into the event relation learner, and use the event relation learner to learn each adjacent event relation alternately to train the event relation extraction process;

[0094] S7: Based on the event argument extraction process trained in S5 and the event relation extraction process trained in S6, jointly extract event arguments and event relations from the text to be identified.

[0095] In this embodiment, as Figure 2 The detailed flowchart shown illustrates how the BERT model is used to encode the text and obtain its embedding vector, where the text is represented as S = {w1, w2, ..., w...}. n}, w n The text embedding vector represents the nth word in the text as: H = {v1, v2, ..., v...} n}, v n Let H represent the embedding vector of the nth word, where H ∈ R. n×d v i ∈R d .

[0096] like Figure 3 The flowchart shown illustrates the event argument extraction process. Semantic dependency analysis tools are used to perform semantic dependency analysis on the text, obtaining semantic dependency information between words and converting it into a structured form to obtain a semantic dependency graph. The semantic dependency graph is represented as: R = (W, Rela), W = {m1, m2, ..., m...} n} represents a set of nodes, and Rela represents a set of semantic dependencies.

[0097] Specifically, semantic dependency analysis tools that can be used include HanLP, DDParser, and OpenHowNet, which perform semantic structure analysis on text to reveal the semantic relationships between words.

[0098] To distinguish between event arguments and non-event arguments and to simplify operations, semantically dependent edges are simplified, resulting in a simplified text graph. The simplified text graph edges defined include: Trigger-Entity edges, Entity-Entity-e edges, Entity-Entity-n edges, and Trigger-Trigger edges. Different types of edges represent relationships between different words, and each edge has a unique identifier, expressed as:

[0099]

[0100] In the formula, ID edge The edge number is indicated by: Trigger-Triggeredge indicates that both ends of the edge are trigger words; Trigger-Entityedge indicates that both ends of the edge are trigger words and words connected to the trigger words; Entity-Entity-e edge indicates that both ends of the edge are event arguments; and Entity-Entity-n edge indicates that both ends of the edge are event arguments and non-event arguments.

[0101] Construct a grid T of size n×n. For each element T in T... i,j That is, the relationship between the i-th word and the j-th word, which is filled using the edge numbers of the corresponding types, expressed as:

[0102]

[0103] In the formula, T i,j This indicates the relationship between the i-th word and the j-th word. A label indicating the relationship between the i-th word and the j-th word;

[0104] After filling, a simplified text graph G = (W, E) is obtained, where W represents the set of nodes in the simplified text graph, the nodes represent words in the text, and E represents the set of edges in the simplified text graph, that is, the elements of the grid T, and the relationships between words.

[0105] For example, the text is "Urea: As the weather warms, the demand for fertilizer increases for spring plowing, wheat regrowth in the north, and rice cultivation in the south, leading to a steady rise in prices." Using a semantic dependency analysis tool, a semantic dependency graph of this text can be obtained. In the semantic dependency graph, there are various types of edges between nodes. To simplify subsequent operations, these edges are converted into the four types of edges defined in this invention. According to the semantic dependency graph, there are three trigger words: "warming," "increase," and "rise." First, words directly connected to the trigger words are identified, such as "warming" connected to "weather," "rise" connected to "price," and "increase" connected to "demand" and "urea." In the text graph, the edges between these words are converted into Trigger-Entity edges. Second, other words connected to the same trigger word are connected. The words "demand" and "urea" are both connected to "increase." In the text graph, the edges between them are defined as Entity-Entity-e edges. "As" and "weather" are both associated with "warming up," and "price" and "stable" are both associated with "rising." However, since "as" and "stable" cannot serve as entity arguments representing any event, these two pairs cannot be connected by an Entity-Entity-e edge. Then, after identifying the trigger words and event arguments, the remaining words are connected to the event arguments using Entity-Entity-n edges. Finally, the trigger words appearing in the text are connected. In this example, there is a relationship between "increase" and "rising," and between "warming up" and "increase," with edges between them being Trigger-Trigger edges.

[0106] In this embodiment, the simplified text graph is divided based on metapaths to obtain metapath subgraphs. These metapath subgraphs include: ETE metapath subgraph, EOE metapath subgraph, and ETO metapath subgraph.

[0107] The simplified text graph is divided according to the ETE path to obtain the ETE meta-path subgraph; the ETE path is defined by the nodes on the path being event arguments, trigger words, and event arguments in sequence.

[0108] The simplified text graph is divided according to the EOE path to obtain the EOE meta-path subgraph; the EOE path is defined by the nodes on the path being event arguments, non-event arguments, and event arguments in sequence.

[0109] The simplified text graph is divided according to the ETO path to obtain the ETO meta-path subgraph; the ETO path is defined as the nodes on the path being event arguments, trigger words and non-event arguments in sequence.

[0110] ETE path partitions the text graph, finds the trigger word node in the text, and searches for the next node that is an event argument. If the next node is an event argument, it is added to the subgraph. Two different event arguments are selected and combined with the trigger word to form a combination of event argument, trigger word, and event argument, thus forming an ETE metapath subgraph, represented as follows. Finally, all of them are combined into a set of ETE metapath subgraphs, represented as follows:

[0111] The EOE path divides the simplified text graph. It selects an event argument node in the text and searches around that node. If a connected node is either an event argument or a non-event argument, it is added to the meta-path. The nodes are combined in the order of event argument-non-event argument-event argument to form a meta-path subgraph, represented as follows: Finally, all possible combinations are synthesized into an EOE metapath subgraph set, represented as follows:

[0112] ETO path divides the simplified text graph. It selects a trigger word node in the text and searches for the next event argument node. If the next node is an event argument, the search continues forward. Then, starting from the event argument node, it searches for non-event argument nodes. The nodes on the path are combined in the order of event argument-trigger word-non-event argument to form a meta-path subgraph, represented as: Finally, all combinations are treated as a set of ETO metapath subgraphs, represented as follows:

[0113] In this embodiment, the process of using a graph convolutional network to perform cross-view comparative learning on the simplified text graph and the meta-path subgraph to obtain node features under the simplified text graph view and the meta-path view respectively is as follows:

[0114] First, construct the adjacency matrix of the simplified text graph; then, select the nodes of the meta-path subgraph and their corresponding edges in the adjacency matrix, and construct the adjacency matrix of the meta-path subgraph based on the selected nodes and their corresponding edges.

[0115] Next, the adjacency matrix of the simplified text graph, the adjacency matrix of the metapath subgraph, and the text embedding vector are input into the graph convolutional network to obtain the node features under the simplified text graph view and the node features under the metapath view, expressed as follows:

[0116] L G =ρ(A G HW0)

[0117] L M =ρ(A M HW0)

[0118] In the formula, LG L represents the node features in a simplified text graph view. M Represents the node features under the meta-path view, ρ(·) represents the activation function, W0 represents the weight matrix, and A G Let A represent the adjacency matrix of the simplified text graph. M H represents the adjacency matrix of the metapath subgraph, and H represents the text embedding vector.

[0119] In this embodiment, the node features in the simplified text graph view and the node features in the meta-path view are mapped to the same feature space using an MLP multilayer perceptron, resulting in node features after MLP transformation, expressed as:

[0120]

[0121] In the formula, This represents the text graph node features after MLP transformation. W represents the node features of the metapath subgraph after MLP transformation. 1 Let W represent the weight matrix. 2 Let b represent the weight matrix. 1 b represents the bias term. 2 This indicates the bias term.

[0122] In this embodiment, the output of the MLP (Multilayer Perceptron) is converted into a label probability and input into the classification layer to obtain the predicted label probability p∈R for each word. n×2 Then, the event arguments are determined based on the predicted labels, expressed as:

[0123] p = softmax(ρ(W) c L+b c ))

[0124] In the formula, ρ represents the nonlinear activation function, and W c Let b represent the weight matrix. c represents the bias term, softmax represents the activation function, and L represents the input to the classification layer.

[0125] In this embodiment, as Figure 4 The flowchart shown illustrates the event relationship extraction process. An event relationship learner is used to learn each type of event relationship alternately, acquiring features for each relationship. The learner transforms the learning task for each event relationship into a binary classification task, i.e., determining whether such a relationship exists. In this embodiment, the text embedding vector is input to the event relationship learner, which then learns each adjacent event relationship alternately. The expression is:

[0126]

[0127] In the formula, σ(·) represents the activation function, W represents the weight matrix, and b represents the bias term. H represents the predicted result of the event relationship output, where H represents the text embedding vector.

[0128] Specifically, the event relationship learner can be composed of a multilayer perceptron (MLP). The text embedding vector is input into the MLP, the event relationship encoder extracts the features related to the event relationship from the text, the event relationship is then classified by the classification layer, and finally the predicted result of the event relationship is output.

[0129] Example 2

[0130] This embodiment provides a detailed explanation of the steps involved in the process of extracting event arguments and training the process of extracting event relationships.

[0131] In the training process for event argument extraction, node features in the meta-path view mapped to the same feature space are used as target features, and node features in the simplified text graph view are used as contrastive features. Contrast loss is calculated and used as the loss function for the graph convolutional network. In contrastive learning, the similarity between positive samples is maximized, while the similarity between positive and negative samples is minimized. After obtaining the output of a sample pair, the similarity between the two outputs is calculated, and the loss value is calculated based on the similarity, thereby optimizing the event argument extraction process. In this embodiment, the contrastive loss is divided into three parts: ETE path loss, EOE path loss, and ETO path loss.

[0132] The ETE path loss is calculated by treating the node features of the ETE meta-path subgraph as positive samples and the features of the remaining meta-path subgraphs as negative samples. The expression is as follows:

[0133]

[0134] In the formula, SG represents the ETE path loss. ETO Represents the ETO metapath subgraph, SG ET1 Metapath Subgraph (ETE), SG EoE denoted as EOE metapath subgraph, exp represents the exponential function, and sim represents the similarity. Indicates positive sample pairs in the output. and similarity, express The sum of similarities with all negative samples.

[0135] The expression for similarity is:

[0136]

[0137] EOE path loss means that the node features of the ETE meta-path subgraph are used as positive samples, and the features of the remaining meta-path subgraphs are used as negative samples to calculate the path loss. The expression is:

[0138]

[0139] The ETO path loss is calculated by treating the node features of the ETO meta-path subgraph as positive samples and the features of the remaining meta-path subgraphs as negative samples. The expression is as follows:

[0140]

[0141] During the training of the event argument extraction process, the contrastive loss is calculated and used as the loss function of the graph convolutional network. The expression is as follows:

[0142]

[0143] In the formula, Represents the loss function. Indicates ETE path loss, This represents the EOE path loss. This represents the ETO path loss.

[0144] In this embodiment, when training the event relationship extraction process, a loss function for calculating the event relationship learner is constructed, and the parameters of the event relationship learner are optimized based on the loss function;

[0145] The loss function of the event relation learner is calculated as follows:

[0146]

[0147] In the formula, This represents the loss function of the event-relationship learner. y represents the predicted result of the event relationship in the output. i Indicates the true outcome of the relationship between events;

[0148] The expression for the parameter optimization process is:

[0149]

[0150] In the formula, θt represents the parameters of the optimized event relation learner, θt-1 represents the parameters of the current event relation learner, and γ represents the learning rate. This represents the gradient of the loss function with respect to the parameters.

[0151] When learning different event relationships, the learner optimizes gradients in different directions, which may affect the learning result of the previous event relationship. To avoid confusion in gradient directions, it is necessary to adjust the gradient direction to reduce repetition with the previous gradient direction. When the learner learns the i-th event relationship, the gradient obtained is... When learning the (i+1)th event relationship, the gradient obtained is The adjustment expression is:

[0152]

[0153] In the formula, gradient and The dot product of two gradient vectors represents the similarity between them. If the two gradient vectors are orthogonal, this value is 0. Gradient In gradient The projection vector in the direction, through the above formula, is the gradient. minus The gradient component in the direction, the remaining component is the same as the gradient component in the direction. Orthogonal gradient components. Changing the direction of the gradient vector can reduce interference between different learning tasks.

[0154] Example 3

[0155] This embodiment provides an event extraction framework based on semantic dependency analysis. This framework is used to implement the event argument extraction process and the event relation extraction process, such as... Figure 5 The framework diagram shown includes: an event argument extraction subframe and an event relation extraction subframe;

[0156] The event argument extraction sub-framework includes: semantic dependency analysis structure, text graph construction structure, meta-path subgraph structure, graph convolutional network structure, and MLP structure;

[0157] The event relationship extraction sub-framework includes a text encoder structure and an event relationship learner structure.

[0158] Example 4

[0159] This embodiment provides an event extraction system based on semantic dependency analysis, such as... Figure 6 The diagram shown illustrates the structure of the system, which includes:

[0160] The text encoding module is used to encode text and obtain the text embedding vector;

[0161] A simplified text graph construction module is used to perform semantic dependency analysis on text. Based on the semantic dependency analysis results, a text semantic dependency graph structure is constructed. The text semantic dependency graph structure includes nodes and semantic dependency edges. The semantic dependency edges are simplified to obtain a simplified text graph.

[0162] The metapath subgraph construction module is used to divide the simplified text graph based on metapaths to obtain metapath subgraphs;

[0163] The node feature output module is used to perform cross-view comparative learning on the simplified text graph and the meta-path subgraph using a graph convolutional network, respectively, to obtain node features under the simplified text graph view and node features under the meta-path view.

[0164] The event argument extraction process training module is used to map the node features under the simplified text graph view and the node features under the meta-path view to the same feature space, calculate the contrastive loss, and use the contrastive loss as the loss function for training the event argument extraction process.

[0165] The event relationship extraction and training module is used to input the text embedding vector into the event relationship learner, and use the event relationship learner to learn each adjacent event relationship alternately to train the event relationship extraction process;

[0166] The joint extraction module is used to jointly extract event arguments and event relationships from the text to be identified based on the trained event argument extraction process and the trained event relationship extraction process.

[0167] The embodiments described are merely examples to clearly illustrate the present invention and are not intended to limit the implementation of the invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all possible implementations. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for event extraction based on semantic dependency analysis, characterized in that, The method comprises the following steps: S1: encoding the text to obtain an embedding vector of the text; S2: performing semantic dependency analysis on the text, constructing a text semantic dependency graph structure according to the semantic dependency analysis result, the text semantic dependency graph structure comprising nodes and semantic dependency edges, simplifying the semantic dependency edges to obtain a simplified text graph; The semantic dependency analysis tool is used to analyze the semantics of the text, obtain a semantic dependency analysis result, and construct a text semantic dependency graph structure as wherein, denotes a node set, denotes a set of semantic dependency relationships; Different types of edges represent different relationships between words, and each edge has a unique number to represent the expression: In the formula, denotes the number of edges, Trigger-Trigger denotes that both of the words connected by the edge are trigger words, denotes that the words connected by the edge are a trigger word and a word connected to the trigger word, denotes that both of the words connected by the edge are event arguments, denotes that the words connected by the edge are an event argument and a non-event argument; A grid of size is constructed , for the elements in the matrix , i.e. the relationship between the th word and the th word, is filled in using the number of the corresponding type of edge, expression: wherein represents a relationship between the th word and the th word, represents a label of a relationship between the th word and the th word; After the filling is complete, a simplified text graph is obtained wherein represents a set of simplified text graph nodes, represents a set of simplified text graph edges, i.e. the elements of the grid . S3: dividing the simplified text graph based on the meta-path to obtain a meta-path subgraph; S4: performing cross-view contrast learning on the simplified text graph and the meta-path subgraph respectively using a graph convolution network to obtain node features in the simplified text graph view and node features in the meta-path view respectively; S5: mapping the node features in the simplified text graph view and the node features in the meta-path view to the same feature space, calculating a contrast loss, and taking the contrast loss as a loss function for training the event argument extraction process to train the event argument extraction process; S6: inputting the text embedding vector into an event relation learner, and alternately learning each adjacent event relation using the event relation learner to train the event relation extraction process; S7: based on the trained event argument extraction process in S5 and the trained event relation extraction process in S6, jointly extracting event arguments and event relations from the text to be recognized.

2. The event extraction method based on semantic dependency analysis according to claim 1, wherein, The text is encoded using the BERT model to obtain its embedding vector, where the text is represented as... , Indicates the first in the text n For each word, the text embedding vector is represented as: , Indicates the first n Embedding vectors of words. 3.The event extraction method based on semantic dependency analysis of claim 1, wherein, The meta-path subgraph comprises an ETE meta-path subgraph, an EOE meta-path subgraph and an ETO meta-path subgraph; The simplified text graph is divided according to an ETE path to obtain an ETE meta-path subgraph The ETE path is a path in which nodes on the path are an event argument, a trigger word and an event argument in sequence. The simplified text graph is divided according to the EOE path to obtain an EOE meta-path subgraph The EOE path is a path in which nodes on the path are an event theory element, a non-event theory element and an event theory element in turn. According to the ETO path, the simplified text graph is divided to obtain an ETO meta path subgraph ; the ETO path is a path in which nodes on the path are an event argument, a trigger word and a non-event argument in turn Finally, a set of subgraphs based on meta-paths is obtained .

4. The event extraction method based on semantic dependency analysis according to claim 1, wherein, The process of performing cross-view contrast learning on the simplified text graph and the meta-path subgraph respectively using a graph convolution network to obtain node features in the simplified text graph view and node features in the meta-path view is as follows: First, construct an adjacency matrix of the simplified text graph; filter out the nodes of the meta-path subgraph and the corresponding edges in the adjacency matrix, and take the filtered nodes and the corresponding edges as the adjacency matrix of the meta-path subgraph; Next, input the adjacency matrix of the simplified text graph, the adjacency matrix of the meta-path subgraph and the text embedding vector into the graph convolution network respectively to obtain the node features in the simplified text graph view and the node features in the meta-path view, the expression being: wherein, denotes the node features under the simplified text graph view, denotes the node features under the meta-path view, denotes the activation function, denotes the weight matrix, denotes the adjacency matrix of the simplified text graph, denotes the adjacency matrix of the meta-path subgraph, H denotes the text embedding vector.

5. The event extraction method based on semantic dependency analysis according to claim 4, characterized in that, The node features under the simplified text graph view and the node features under the meta-path view are mapped to the same feature space by using an MLP to obtain and , and the expressions are respectively In the formula, denotes the text graph node feature converted through the MLP, denotes the meta-path subgraph node feature converted through the MLP, denotes a weight matrix, denotes a weight matrix, denotes a bias term, denotes a bias term; Take the node features in the meta-path view mapped to the same feature space as the target features, take the node features in the simplified text graph view as the contrast features, calculate the contrast loss, and take the contrast loss as the loss function of the graph convolution network; The contrast loss is divided into three parts, which are ETE path loss, EOE path loss and ETO path loss; The ETE path loss represents that the node features of the ETE meta-path subgraph are taken as positive samples, and the node features of the remaining meta-path subgraphs are taken as negative samples, and the path loss is calculated, the expression being: wherein, denotes the ETE path loss, denotes the ETO meta-path subgraph, meta-path subgraph ETE meta-path subgraph, denotes the EO E meta-path subgraph, exp denotes the exponential function, denotes the similarity; denotes the similarity of the positive sample pair output and denotes the sum of the similarities with all negative samples;​ The EOE path loss represents that the node features of the ETE meta-path subgraph are taken as positive samples, and the node features of the remaining meta-path subgraphs are taken as negative samples, and the path loss is calculated, the expression being: In the formula, denotes the EOE path loss; The ETO path loss represents that the node features of the ETO meta-path subgraph are taken as positive samples, and the node features of the remaining meta-path subgraphs are taken as negative samples, and the path loss is calculated, the expression being: The expression of the contrast loss is: wherein denotes the contrast loss, denotes the ETE path loss, denotes the EOE path loss, denotes the ETO path loss.

6. The event extraction method based on semantic dependency analysis according to claim 1, wherein, The text embedding vector is input into an event relation learner constructed by an MLP, and each adjacent event relation is alternately learned by using the event relation learner; Event relation learner for the first i Learn the relationships between events to obtain output. The expression is: wherein, denotes an activation function, denotes a weight matrix, denotes a bias term, denotes the predicted result of the output event relation, denotes a text embedding vector.

7. The event extraction method based on semantic dependency analysis according to claim 6, wherein, In the training of the event relation extraction process, a loss function for calculating the event relation learner is constructed, and the parameters of the event relation learner are optimized according to the loss function; The calculation expression of the loss function of the event relation learner is: In the formula, denotes a loss function of the event relation learner, denotes a prediction result of the output event relation, denotes a true result of the event relation; The expression of the parameter optimization process is: wherein θ t denotes the parameters of the optimized event relation learner, θ t 1denotes the parameters of the current event relation learner, γ denotes the learning rate, denotes the gradient of the loss function with respect to the parameters.

8. An event extraction framework based on semantic dependency analysis, characterized in that, The event extraction framework is used to implement the event argument extraction process and the event relation extraction process in claim 1, and includes an event argument extraction sub-framework and an event relation extraction sub-framework. The event argument extraction sub-framework includes a semantic dependency analysis structure, a text graph construction structure, a meta-path sub-graph structure, a graph convolution network structure and an MLP structure. The event relation extraction sub-framework includes a text encoder structure and an event relation learner structure.

9. A system for event extraction based on semantic dependency analysis, for implementing the method for event extraction based on semantic dependency analysis as claimed in any one of claims 1 to 7, characterized in that, The text encoding module is configured to encode the text to obtain a text embedding vector. The simplified text graph construction module is configured to analyze the semantic dependency of the text, construct a text semantic dependency graph structure according to the semantic dependency analysis result, and obtain a simplified text graph including nodes and semantic dependency edges and simplified semantic dependency edges. The meta-path sub-graph construction module is configured to divide the simplified text graph based on a meta-path to obtain a meta-path sub-graph. The node feature output module is configured to perform cross-view contrast learning on the simplified text graph and the meta-path sub-graph by using a graph convolution network to obtain node features in a simplified text graph view and node features in a meta-path view, respectively. The event argument extraction process training module is configured to map the node features in the simplified text graph view and the node features in the meta-path view to the same feature space, calculate a contrast loss, and use the contrast loss as a loss function for training the event argument extraction process to train the event argument extraction process. The event relation extraction and training module is configured to input the text embedding vector into the event relation learner, and alternately learn each adjacent event relation by using the event relation learner to train the event relation extraction process. The joint extraction module is configured to jointly extract event arguments and event relations from the to-be-identified text based on the trained event argument extraction process and the trained event relation extraction process. ​

Images