Event causal relationship identification method and system based on dual-channel and multilayer graph convolutional network

Abstract

The invention discloses an event causal relationship identification method and system based on two channels and a multilayer graph convolutional network, and relates to natural language processing. Comprising the following steps: inputting and preprocessing a text to obtain a lexical sequence, event mention and an event set; performing dual-channel coding, and fusing BiLSTM (Bidirectional Long Short Term Memory) and Transform to extract local and global features; performing multi-layer graph convolutional network processing, constructing a lexical-level, mention-level and event-level three-layer graph structure, and aggregating information through a GCN with residual connection; predicting a causal relationship, fusing multi-source features to construct event embedding, and predicting the causal relationship through an MLP classifier; and training and optimizing the model. The system comprises text input, two-channel coding, a multilayer graph convolutional network, a prediction module and the like. The method solves the problems of insufficient co-reference relationship modeling, weak hierarchical semantic capture and the like, improves the recognition precision of the causal relationship of the document-level event, has high generalization ability and practical level, and can be widely applied to the scenes of public opinion analysis, disaster detection and the like.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for identifying event causal relationships based on dual-channel and multi-layer graph convolutional networks. Background Technology

[0002] Event Causality Identification (ECI) is a core task in Natural Language Processing (NLP), aiming to identify causal relationships between events in text. This technology addresses the ambiguity, implicitness, and polysemy of causal expressions in natural language. By analyzing the trigger word features, semantic structure, and contextual logical connections of events, it can determine whether a causal relationship exists between two or more events. In applications, ECI technology can extract deep causal logic from unstructured data such as news corpora, medical records, and legal documents, significantly improving machine learning's ability to understand and reason about text semantics, and providing reliable evidence for scenarios such as public opinion analysis, disaster detection, and fault diagnosis.

[0003] Graph Convolutional Networks (GCNs), as an important branch of graph neural networks, have demonstrated unique advantages in processing non-Euclidean data structures in recent years. GCNs update the representation of the current node by aggregating information from neighboring nodes, capturing both local and global structural features in the graph, thus effectively handling complex relationships between nodes. This technology effectively overcomes the technical bottleneck of traditional convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in handling irregular graph data, and can efficiently model complex graph structures composed of nodes and edges. In the field of natural language processing, GCNs are widely used in tasks such as semantic annotation, relation extraction, and text classification. By modeling words and events as graph nodes and constructing edge relationships using syntactic dependencies and semantic associations, GCNs can effectively integrate multi-source language information and improve the model's semantic understanding capabilities.

[0004] Despite the progress made by existing technologies in ECI tasks and GCN applications, current event causality identification technologies still have the following shortcomings, mainly:

[0005] 1. The same event often appears multiple times in a document using different vocabulary, forming complex coreference relationships. Traditional methods typically treat events as isolated semantic units, failing to adequately model the mention relationships between events. This makes it difficult for the model to capture semantic dependencies across sentences, limiting its ability to understand document-level causal logic.

[0006] 2. Most existing graph neural network-based methods use a single heterogeneous graph structure, which mixes nodes of different semantic levels (such as lexical, mention, and event) in the same graph. This makes it impossible to capture the interaction relationships between nodes within the same level, thus limiting the model's ability to perceive and reason about the hierarchical semantic structure of documents.

[0007] 3. In news analysis scenarios, the lack of system applications that support real-time text analysis, cause-effect graph visualization, and user interaction limits the application value of ECI technology in practical scenarios such as public opinion analysis and event prediction.

[0008] Therefore, there is an urgent need for an event causal relationship recognition technology that can effectively model hierarchical semantic associations, integrate multi-dimensional features, and has universal feasibility, in order to solve the defects and shortcomings of existing technologies. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies, such as insufficient modeling of coreference relationships, weak hierarchical semantic capture, and low practicality, by providing an event causal relationship recognition method and system based on dual-channel and multi-layer graph convolutional networks. This method and system can achieve accurate recognition of document-level event causal relationships while improving the practicality of the technology.

[0010] To achieve the above-mentioned objectives, the present invention provides the following technical solutions.

[0011] An event causal relationship identification method based on dual-channel and multi-layer graph convolutional networks includes the following steps:

[0012] 1) Text input and preprocessing: Obtain the text of the document to be processed, perform word segmentation, part-of-speech tagging, syntactic analysis and event extraction on the text to obtain word sequence, syntactic dependency relationship, event mention set and event set;

[0013] 2) Dual-channel encoding: Based on the word sequence obtained in step 1), construct a word input vector that integrates word embedding, event embedding and mention embedding, and input it in parallel into the BiLSTM encoder and the Transformer encoder to extract the local context feature sequence and the global semantic representation sequence, respectively;

[0014] 3) Multi-layer graph convolutional network processing: Based on the local context feature sequence and global semantic representation sequence output in step 2), word-level, mention-level and event-level graph structures are constructed sequentially. The feature sequence is used as the initial node features of each graph structure. The information of each graph structure is aggregated step by step through a multi-layer graph convolutional network (GCN) with residual connections to output event-level node representations.

[0015] 4) Causal relationship prediction: The event-level node representation output from step 3), the global semantic representation output from step 2), and the word-level representation after mapping transformation are fused to construct an event fusion embedding; the fusion embedding of any event pair is concatenated and input into a multilayer perceptron (MLP) classifier to predict whether there is a causal relationship between the event pairs;

[0016] 5) Model training and optimization: The above model is trained based on the labeled dataset. The model parameters are optimized by minimizing the cross-entropy loss function, and a regularization mechanism is introduced to prevent overfitting.

[0017] In step 1), the event extraction adopts a method based on trigger word identification and argument extraction. Trigger word identification uses the pre-trained language model BERT-base to classify and determine words such as verbs and nouns in the text, and selects trigger words with event triggering capabilities. Based on syntactic dependencies, the core arguments corresponding to each trigger word are determined, including agent, patient, time, place, etc., and a complete event mention is composed of "trigger word + core argument". Core reference resolution is performed on all event mentions, and multiple mentions pointing to the same real event are grouped into one category to form an event set.

[0018] In step 2), the specific steps of the dual-channel encoding can be as follows:

[0019] 2.1) Construct an input vector that fuses information from multiple sources. For the first element in the document... Each word element has an input representation. It is composed of word embeddings, event embeddings, and mention embeddings, and the calculation formula is as follows:

[0020]

[0021] in, , and These represent the numbers for words, events, and mentions, respectively. , and This corresponds to the embedding layer; for word embeddings Pre-trained language model (BERT) and static word vectors (GloVe) are used for initialization to introduce prior lexical semantic knowledge; the entire document's word input vector sequence... They are then fed into two independent encoding channels in parallel for processing;

[0022] 2.2) The first channel uses a bidirectional long short-term memory network (BiLSTM) to capture the local contextual semantic features of lexical units; for each lexical unit position... BiLSTM calculates its forward hidden state respectively. and backward hidden state The computation process follows the recursive update mechanism of LSTM:

[0023]

[0024]

[0025] The output of this channel is the first Local context representation of each word It is formed by splicing together its forward and backward hidden states:

[0026]

[0027] The second channel employs a Transformer encoder, utilizing a self-attention mechanism to model the global contextual semantic features of tokens; the self-attention layer calculates attention weights between token pairs based on the query, key, and value matrix; for the input sequence Its self-attention output The calculation is as follows:

[0028]

[0029] in, It is a parameter matrix. It is the hidden layer dimension. It is the scaling factor;

[0030] 2.3) Self-attention output A feedforward neural network (FFN) is used, which includes two linear transformation layers and a non-linear activation function (ReLU) to further enhance the representation capability.

[0031]

[0032] in, Represents the weight matrix. Represents the bias vector; The global context representation of the output of this channel;

[0033] 2.4) For each token, the dual-channel encoding module outputs two complementary representation sequences: a local context representation sequence generated by the BiLSTM channels. and the global semantic representation sequence generated by the Transformer channel. These two sequences will serve as the initial features of nodes in different levels of graph structures in subsequent multi-layer graph convolutional network modules, laying the foundation for hierarchical information aggregation in the model.

[0034] In step 3), the specific steps of the multi-layer graph convolutional network processing include:

[0035] 3.1) In a lexical graph, nodes are all the lexical terms in a document; seven types of edges are defined to construct the graph's topology: syntactic dependency edges (connecting based on syntactic rules), mention-within edges (connecting all lexical terms within the same event mention), mention-within edges (connecting lexical terms between mentions of different events), event edges (connecting lexical terms of different events), sentence-within edges (connecting adjacent lexical terms within the same sentence), sentence-within edges (connecting the first and last lexical terms of adjacent sentences), and self-loop edges (each node connects to itself).

[0036] Lexical-level graphs are sequences of local context representations output by the previous module. As the initial features of the nodes, a multi-layer graph convolutional network is then used to learn the graph; for the th The update formula for nodes in a layered graph convolutional network is as follows:

[0037]

[0038]

[0039] in, It is the adjacency matrix of the word-level graph. and It is the first Layer parameters, It's an activation function; here, a residual connection is introduced, concatenating the outputs of the first two layers as a supplement to the input of the current layer. To mitigate the oversmoothing problem that may occur in deep GCNs, such as the phenomenon where node features tend to be consistent during the training of deep GCN models, leading to a loss of discriminative ability; the output of the word-level graph convolutional network is the word representation after deep aggregation. ;

[0040] 3.2) In the mention-level graph, nodes are event mentions; they mainly contain three types of edges: mention-mentioned edges (connecting mentions pointing to the same event), event-event edges (connecting mentions of different events in the same sentence), and self-looping edges; these are mapped through a lexical matrix. lexical representation and global semantic representation Perform transformation and fusion to initialize the representation of each event mention node:

[0041]

[0042]

[0043]

[0044] Initial node representation The information is fed into a mention-level graph convolutional network for propagation and aggregation. Its inter-layer update method is similar to that of a word-level graph convolutional network, using an adjacency matrix. This allows us to aggregate information mentioned by neighbors, thereby obtaining a mention representation that contains document-level information. ;

[0045] 3.3) In an event-level graph, nodes represent events within a document. First, the central event in the document is determined based on its frequency of mention, and it is connected to all other event nodes. Furthermore, co-occurrence edges are established between event pairs mentioned in the same sentence. The initial representation of event nodes... By mentioning the mapping matrix Reference level and Obtained through fusion:

[0046]

[0047]

[0048]

[0049] This initial representation It is then fed into an event-level graph convolutional network, based on the adjacency matrix of the event-level graph. The convolution operation is performed, and the final output is an event node representation that reflects the potential causal relationship between events and the topic association. ;

[0050] By sequentially processing and transmitting information through three hierarchical graph structures, a step-by-step abstraction and reasoning process is achieved, from lexical and grammatical information to document themes.

[0051] In step 4), the specific steps for predicting the causal relationship can be as follows:

[0052] 4.1) Representing by piecing together event-level diagrams Global context representation and word-level information after mapping transformation The feature vector for each event Construct a fused embedded representation:

[0053]

[0054] in, yes The corresponding representation, yes The corresponding representation, yes The corresponding representation ensures that the event representation simultaneously contains local details, global context, and structured relationships;

[0055] 4.2) For any pair of events that needs to be determined, the module will embed the final representation of the two events. and Perform vector concatenation to form a joint representation of the event pair. :

[0056]

[0057] in, This represents a vector concatenation operation; It was then fed into a multilayer perceptron (MLP) for causal classification;

[0058] 4.3) The MLP performs a series of nonlinear transformations to map the high-dimensional features of the input to the probability space of causal categories, and outputs the probability distribution of each causal category through the Softmax activation function:

[0059]

[0060] in, and These are the weight matrix and bias vector of the last layer in an MLP classifier. It is a two-dimensional probability vector, representing the predicted probabilities of "a causal relationship exists" and "a causal relationship does not exist" between event pairs, respectively;

[0061] 4.4) Model training is performed at the document level; for a single document, the loss function is defined as... The sum of the cross-entropy losses of each event pair is used, and an L2 regularization term is introduced to control model complexity and prevent overfitting.

[0062]

[0063] in, Representing the The true causal relationship label for each event pair. It is the probability distribution predicted by the model. It is the regularization coefficient. This represents all trainable parameters of the model.

[0064] An event causal relationship recognition system based on dual-channel and multi-layer graph convolutional networks, comprising a text input module, a dual-channel encoding module, a multi-layer graph convolutional network module, and a prediction module.

[0065] The text input module is used to receive the text of the document to be processed, perform word segmentation and part-of-speech tagging preprocessing, and output a word sequence.

[0066] The dual-channel encoding module is used to construct multi-source fused word input vectors, and extracts local contextual features and global semantic representations in parallel using BiLSTM and Transformer.

[0067] The multi-layer graph convolutional network module is used to construct a three-layer graph structure at the word level, mention level, and event level. It aggregates information from each layer through a multi-layer GCN and outputs an event-level representation.

[0068] The prediction module is used to fuse multi-source features to construct event embeddings and output causal relationship prediction results through an MLP classifier.

[0069] Furthermore, the text input module is implemented using the Python programming language combined with the NLTK toolkit. It is used to receive the text of the document to be processed (supporting txt, docx, and pdf formats), and perform word segmentation, part-of-speech tagging, syntactic analysis, and event extraction operations. Specifically, word segmentation uses NLTK's WordPunctTokenizer, part-of-speech tagging uses NLTK's PerceptronTagger, syntactic analysis uses spaCy, and event extraction is implemented based on a pre-trained BERT-base model. The output includes a standardized lexical sequence, a set of syntactic dependencies, a set of event mentions, and a set of events.

[0070] Furthermore, the dual-channel encoding module includes a multi-source embedding unit, a BiLSTM encoding unit, and a Transformer encoding unit; the multi-source embedding unit is used to construct a lexical input vector that integrates word embeddings, event embeddings, and mention embeddings; the BiLSTM encoding unit is used to extract the local contextual feature sequence of lexical units; and the Transformer encoding unit is used to extract the global semantic representation sequence of lexical units.

[0071] Furthermore, the multi-layer graph convolutional network module includes a word-level graph construction unit, a mention-level graph construction unit, an event-level graph construction unit, and a multi-layer GCN unit; the word-level graph construction unit is used to construct a word-level graph with seven edge types; the mention-level graph construction unit is used to construct a mention-level graph with three edge types; the event-level graph construction unit is used to construct an event-level graph with edges associated with a central event; and the multi-layer GCN unit is used to aggregate information from each layer graph structure and output an event-level representation.

[0072] Furthermore, the prediction module includes a feature fusion unit and an MLP classification unit; the feature fusion unit is used to concatenate event-level representation, global semantic representation and transformed word-level representation to obtain event fusion embedding; the MLP classification unit is used to classify the event pair fusion embedding and output causal relationship prediction results.

[0073] Furthermore, the system may also include a visualization interaction module, which is used to realize text input interaction, display of causal relationship results, and interactive causal graph visualization functions. The visualization interaction module can be built using general Web development technologies and the Streamlit 1.44.0 framework and connected to the prediction module. The core functions include: ① Text input interface: supports batch / single text input; ② Result display: displays the extracted events and identified causal pairs in tabular form; ③ Interactive causal graph: the node size and color depth map the frequency of causal associations of events, supports zooming, dragging, and viewing node details, and edges represent causal relationships.

[0074] Furthermore, the visualization interaction module is built on a Web platform based on Streamlit, supporting text input, one-click analysis, table display, and cause-effect graph zooming and drag-and-drop interaction.

[0075] The event causal relationship recognition method proposed in this invention, based on dual-channel and multi-layer graph convolutional networks, achieves significant technical progress in document-level event causal relationship recognition tasks by introducing a multi-layer graph convolutional network structure and a dual-channel feature encoding mechanism. Compared with existing technologies, the outstanding technical effects and advantages of this invention are as follows:

[0076] 1. This invention utilizes a dual-channel encoding module to parallelly fuse the local feature capture capability of BiLSTM with the global semantic modeling capability of Transformer, providing robust initial node features for subsequent graph reasoning. Simultaneously, it constructs a three-layer graph structure at the word level, mention level, and event level, achieving cross-level and cross-sentence information aggregation through multi-layer GCNs, effectively addressing the problems of insufficient coreference modeling and weak hierarchical semantic capture. Experiments on the EventStoryLine v1.5 dataset show that the method of this invention achieves an overall F1 score of 60.8% and a cross-sentence causal recognition F1 score of 54.5%, significantly outperforming baseline methods such as ILP, RichGCN, and ERGO.

[0077] 2. This invention introduces prior lexical semantic knowledge by jointly initializing word embeddings using BERT and GloVe; it uses residual connections in each layer of GCN to alleviate the oversmoothing problem, and combines Dropout and L2 regularization to control model complexity, so that the model has stable performance in causal recognition tasks of different types of documents (such as news and medical records).

[0078] 3. The system of this invention can be implemented using a universal technology stack (Python, PyTorch, Streamlit, etc.), supporting fully automated processing from text input, event extraction, cause-effect recognition to cause-effect graph visualization; it supports interactive operation, and the web-based interactive interface is simple and easy to use, and can be directly applied to practical scenarios such as public opinion analysis and emergency management, significantly enhancing the practical value of ECI technology. Attached Figure Description

[0079] Figure 1 This is a diagram showing the overall structure of the DMGCN model of this invention.

[0080] Figure 2 This invention is a word-level word. Figure 7 Example diagram of edge planting.

[0081] Figure 3 This is an overall flowchart of the method of the present invention;

[0082] Figure 4 This is a schematic diagram of the structure of the dual-channel encoding module of the present invention;

[0083] Figure 5 This is a schematic diagram of the structure of the multilayer graph convolutional network module of the present invention;

[0084] Figure 6 This is a module architecture diagram of the system of the present invention;

[0085] Figure 7 This is a schematic diagram of the interface of the visual interaction module of the present invention.

[0086] Figure 8 A causal relationship diagram generated for a specific example of the present invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of this invention clearer, the following embodiments will be used in conjunction with the accompanying drawings to further illustrate the invention. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Rather, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Details not described in detail can be achieved using conventional methods in the art.

[0088] This invention proposes a method and system for event causality recognition based on dual-channel and multi-layer graph convolutional networks. The method integrates three levels of graph structures (term level, mention level, and event level) to model semantic interactions across different levels. By applying graph convolutional networks at each level, rich node information is captured, enhancing implicit causal recognition capabilities. Furthermore, a dual-channel encoding module is designed to effectively fuse global contextual semantic features. A visualization interaction module is also developed, supporting the entire process from text input to causal graph visualization, significantly improving the practicality of event causality recognition technology.

[0089] I. System Overall Architecture

[0090] The overall structure of the DMGCN model of this invention is as follows: Figure 1 As shown, this invention presents an event causal relationship recognition process: "Document Input → Multi-Source Embedding → Dual Encoding Feature Extraction → Three-Layer Graph Convolution Semantic Aggregation → Event Pair Classification → Visualization Output." The invention primarily consists of three core modules: a dual-channel encoding module, a multi-layer graph convolutional network module, a prediction module, and a visualization interaction module. The document text is first input to the dual-channel encoding module, where a BiLSTM encoder and a Transformer encoder are used in parallel to extract local contextual features and global semantic representations. Subsequently, the multi-layer graph convolutional network module constructs graph structures at three levels: word level, mention level, and event level, and applies graph convolutional networks for node representation learning and information aggregation to capture semantic dependencies and implicit causal relationships across sentences. Finally, the prediction module fuses multi-level graph representations and global contextual features, and uses a multi-layer perceptron classifier to determine whether a causal relationship exists between event pairs. Simultaneously, the visualization interaction module, based on the technology of this invention, can intuitively display the visualization effect of the generated causal graph.

[0091] 1.1 Dual-channel encoding module

[0092] The dual-channel encoding module processes the input document text to generate a lexical representation that combines local features and global contextual semantics. By employing two layers of parallel encoding, the vanishing gradient and exploding gradient problems existing in multi-layer networks can be effectively alleviated. Figure 4 This is a schematic diagram of the dual-channel coding module of the present invention, illustrating the collaborative working mechanism of the multi-source embedding unit, the BiLSTM coding unit, and the Transformer coding unit:

[0093] First, construct an input vector that integrates information from multiple sources; for the first... Each word element has an input representation. It is composed of word embeddings, event embeddings, and mention embeddings, and the calculation formula is as follows:

[0094]

[0095] in, , and These represent the numbers for words, events, and mentions, respectively. , and This corresponds to the embedding layer; for word embeddings Pre-trained language model (BERT) and static word vectors (GloVe) are used for initialization to introduce prior lexical semantic knowledge; the entire document's word input vector sequence... They are then fed into two independent encoding channels in parallel for processing.

[0096] The first channel uses a bidirectional long short-term memory (BiLSTM) network to capture the local contextual semantic features of lexical units; for each lexical position... BiLSTM calculates its forward hidden state respectively. and backward hidden state The computation process follows the recursive update mechanism of LSTM:

[0097]

[0098]

[0099] Finally, the output of this channel is the first Local context representation of each word It is formed by splicing together its forward and backward hidden states:

[0100]

[0101] The second channel employs a Transformer encoder, utilizing a self-attention mechanism to model the global contextual semantic features of tokens; the self-attention layer calculates attention weights between token pairs based on the query, key, and value matrix; for the input sequence Its self-attention output The calculation is as follows:

[0102]

[0103] in, It is a parameter matrix. It is the hidden layer dimension. It is the scaling factor.

[0104] Subsequently, It will use a feedforward neural network (FFN) containing two layers of linear transformation and a non-linear activation function (ReLU) to further enhance the representation capability:

[0105]

[0106] in, Represents the weight matrix. Represents the bias vector; The global context representation of the output of this channel.

[0107] Finally, for each token, the dual-channel encoding module outputs two complementary representation sequences: a local context representation sequence generated by the BiLSTM channels. and the global semantic representation sequence generated by the Transformer channel. These two sequences will serve as the initial features of nodes in different levels of graph structures in subsequent multi-layer graph convolutional network modules, laying the foundation for hierarchical information aggregation in the model.

[0108] 1.2 Multilayer Graph Convolutional Network Module

[0109] Compared to traditional sequence models, graph convolutional networks (GCNs) have a natural advantage in handling tasks such as long-distance dependencies and multi-hop reasoning. They can achieve cross-node propagation of information through multi-layer convolutional operations. Multi-layer graph convolutional network modules construct graph structures at three levels: word level, mention level, and event level, and apply graph convolutional networks to aggregate information to capture multi-level semantic information, thereby enhancing the ability to identify implicit causal relationships. Figure 5 The present invention provides a schematic diagram of the structure of the multi-layer graph convolutional network module, which shows the hierarchical relationship and information transmission path of the word-level graph, the mention-level graph, and the event-level graph.

[0110] In a lexical-level graph, nodes represent all the terms in a document; for example... Figure 2 (Word level) Figure 7 As shown in the example of edge type, in order to fully capture the various associations between lexical units, this invention defines seven types of edges to construct the graph topology: syntactic dependency edges (connected based on grammatical rules), mention-within edges (connecting all lexical units within the same event mention), mention-within edges (connecting lexical units between mentions of different events), event edges (connecting lexical units of different events), sentence-within edges (connecting adjacent lexical units in the same sentence), sentence-within edges (connecting the first and last lexical units of adjacent sentences), and self-loop edges (each node connects to itself).

[0111] Lexical-level graphs are sequences of local context representations output by the previous module. As the initial features of the nodes, a multi-layer graph convolutional network is then used to learn the graph. For the ... The update formula for nodes in a layered graph convolutional network is as follows:

[0112]

[0113]

[0114] in, It is the adjacency matrix of the word-level graph. and It is the first Layer parameters, It's an activation function. Here, a residual connection is introduced, concatenating the outputs of the first two layers as a supplement to the input of the current layer. This is to mitigate the oversmoothing problem that may occur in deep GCNs. The output of the lexical-level graph convolutional network is the lexical representation after deep aggregation. .

[0115] In a mention-level graph, nodes are event mentions. There are three main types of edges: mention-mentioned edges (connecting mentions of the same event), event-event edges (connecting mentions of different events within the same sentence), and self-looping edges. This is achieved through a lexical mapping matrix. lexical representation and global semantic representation Perform transformation and fusion to initialize the representation of each event mention node:

[0116]

[0117]

[0118]

[0119] Initial node representation The information is fed into a mention-level graph convolutional network for propagation and aggregation. Its inter-layer update method is similar to that of a word-level graph convolutional network, using an adjacency matrix. This allows us to aggregate information mentioned by neighbors, thereby obtaining a mention representation that contains document-level information. .

[0120] Finally, there's the event-level graph, where nodes represent events within the document. First, the central event in the document is determined based on its mention frequency, and it's connected to all other event nodes. Furthermore, co-occurrence edges are established between event pairs mentioned in the same sentence. The initial representation of event nodes... By mentioning the mapping matrix Reference level and To obtain by fusion:

[0121]

[0122]

[0123]

[0124] This initial representation It is then fed into an event-level graph convolutional network, based on the adjacency matrix of the event-level graph. The convolution operation is performed, and the final output is an event node representation that reflects the potential causal relationship between events and the topic association. .

[0125] Through the sequential processing and information transfer of three hierarchical graph structures, this module achieves layer-by-layer abstraction and reasoning from lexical and grammatical information to document topics, providing powerful feature representations for subsequent causal relationship identification.

[0126] 1.3 Prediction Module

[0127] The prediction module is responsible for determining causal relationships between any event pairs in the document. This module receives event-level graph representations from the multi-layer graph convolutional network module. Global context representation and word-level information after mapping transformation First, the module provides each event with... A fused embedding representation is constructed by concatenating feature vectors from the three sources mentioned above:

[0128]

[0129] in, yes The corresponding representation, yes The corresponding representation, yes The corresponding representation ensures that the event representation simultaneously contains local details, global context, and structured relationships.

[0130] For any pair of events that needs to be determined, the module will embed the final representation of the two events. and Perform vector concatenation to form a joint representation of the event pair. :

[0131]

[0132] in, This indicates a vector concatenation operation. The input features are then fed into a multilayer perceptron (MLP) for causal classification. The MLP performs a series of nonlinear transformations, mapping the high-dimensional features of the input to a probability space of causal categories. Finally, a softmax activation function is used to output the probability distribution of each causal category.

[0133]

[0134] in, and These are the weight matrix and bias vector of the last layer in an MLP classifier. It is a two-dimensional probability vector, representing the predicted probabilities of "a causal relationship exists" and "a causal relationship does not exist" between event pairs.

[0135] Model training takes place at the document level. For a single document, the loss function is defined as follows: The sum of the cross-entropy losses of each event pair is used, and an L2 regularization term is introduced to control model complexity and prevent overfitting.

[0136]

[0137] here, Representing the The true causal relationship label for each event pair. It is the probability distribution predicted by the model. It is the regularization coefficient. This represents all trainable parameters of the model. In this embodiment, the regularization coefficient λ can range from 0.0001 to 0.001, preferably 0.0005; the Dropout probability can be set to 0.3.

[0138] 1.4 Visual Interaction Module

[0139] The visualization and interaction module is responsible for transforming the event causal relationship pairs output by the prediction module into an intuitive and interactive graphical causal graph. Figure 7 Here is an example of the interface of a visual interactive module of the present invention. The interface adopts a left-right split layout: the left side is the function operation area, which includes a "text input box" (supporting batch / single text input), a "one-click analysis" button, an "event extraction result" area, and a "cause-effect event pair display" area; the right side is the visualization area for generating a cause-effect graph. The overall layout is clear.

[0140] To achieve the above functionality, this module is built on the Python programming language and the Streamlit 1.44.0 framework, employing a front-end and back-end separation architecture. Specifically, the front-end layer uses Streamlit components to build an interactive interface, receiving input text and triggering analysis. The back-end layer receives the event list and causal pairs output by the prediction module, calls the PyVis library, calculates node positions using the force-directed layout algorithm, maps events to circular nodes, and maps causal relationships to connecting edges, generating an initial network topology graph. To further enhance the information expression and interactivity of the graph, this module adjusts the size and color depth of nodes based on the number of causal pairs associated with events, thus visually distinguishing the importance of events. Finally, a causal graph supporting zooming, dragging, and node detail viewing is generated and rendered in real-time on the front-end interface, providing a clear presentation of causal relationships.

[0141] II. Model Training

[0142] The model is trained using a supervised learning paradigm, optimizing the loss function to learn the model parameters, enabling it to accurately identify document-level event causal relationships. The entire training process strictly follows the method described in this invention and is performed on the recognized benchmark dataset EventStoryLine v1.5.

[0143] Table 1: Hyperparameter settings for DMGCN

[0144]

[0145] Training was implemented using Python and the PyTorch deep learning framework. For encoding the input sequences, the BERT-based cased pre-trained model from HuggingFace was used. This model contains 12 Transformer layers with a hidden layer dimension of 768. For word embeddings, the model integrated GloVe and BERT embeddings to enhance representation learning. Furthermore, the event embedding dimension and mention embedding dimension were both set to 20. The model was trained using the Adam optimizer with a learning rate of 5e-5 and a fixed batch size of 15. To prevent overfitting, Dropout and L2 regularization were introduced during training. The hidden layer dimension of each layer in the Graph Convolutional Network (GCN) was uniformly set to 100. All trainable model parameters were sampled from a normal distribution during initialization. Specific hyperparameters are shown in Table 1 (DMGCN hyperparameter settings).

[0146] The goal of model training is to minimize the loss function defined by the formula. This function is the sum of the cross-entropy losses of all event pairs, with an additional L2 regularization term. During training, the model undergoes multiple iterations on the training set and its performance is continuously monitored on an independent development set. Finally, the model parameters with the best performance on the development set are saved.

[0147] III. Event Causality Identification Based on the Method of the Invention

[0148] Figure 3 The overall flowchart of the method of this invention is given, including step 1: text input and preprocessing; step 2: dual-channel encoding; step 3: multi-layer graph convolutional network processing; step 4: causal relationship prediction; and step 5: model training and optimization. The following embodiment uses the news text "Due to heavy rainfall causing flooding in many areas, municipal departments immediately activated emergency drainage plans, and the flooding situation gradually eased" as an example to explain in detail the implementation process of the method of this invention:

[0149] Step 1: Text Input and Preprocessing

[0150] Receive the above news text and perform the following operations: ① Word segmentation: Obtain the word sequence [because, heavy rainfall, caused, many, places, flooding,, municipal departments, immediately, activate, emergency drainage plan,, flooding, situation, gradually, alleviate]; ② Part-of-speech tagging: Tag the part of speech of each word (e.g., "heavy rainfall" is a noun, "caused" is a verb, etc.); ③ Syntactic analysis: Obtain syntactic dependencies through spaCy (e.g., "caused" and "heavy rainfall" have a subject-predicate relationship, "caused" and "flooding" have a verb-object relationship, etc.); ④ Event extraction: Identify the trigger words "heavy rainfall", "caused", "flooding", etc. By extracting the core arguments corresponding to each trigger word, we obtain the event mention set: mention 1 (trigger word "heavy rainfall", argument "none"), mention 2 (trigger word "occurrence", argument "flooding"), mention 3 (trigger word "activation", arguments "municipal departments" and "emergency drainage plan"), mention 4 (trigger word "alleviation", argument "flooding situation"). After coreference resolution, the event set is: event A (corresponding to mention 1, heavy rainfall), event B (corresponding to mentions 2 and 4, flooding occurrence and alleviation), and event C (corresponding to mention 3, activation of emergency drainage plan).

[0151] Step 2: Dual-channel encoding

[0152] A multi-source fusion input vector is constructed for each word element. For example, the input vector of the word element "cause" is formed by concatenating its BERT+GloVe word embedding, the embedding of the associated event (if there is no explicit event, the embedding is a zero vector), and the embedding of the associated mention (if there is no explicit mention, the embedding is a zero vector). The input vector sequence is fed into BiLSTM and Transformer in parallel: BiLSTM captures the local dependencies between "cause" and the preceding and following words "heavy rainfall" and "multiple locations", and outputs a local contextual representation; Transformer captures the global associations between "cause" and "heavy rainfall", "flooding", and "alleviation" through a self-attention mechanism, and outputs a global semantic representation.

[0153] Step 3: Multi-layer graph convolutional network processing

[0154] ① Lexical-level graph: Nodes represent all lexical units, constructing seven types of edges, including syntactic dependency edges (e.g., "heavy rainfall" - "caused") and intra-sentence edges (e.g., "caused" - "heavy rainfall"). Local features output by BiLSTM are used as initial features, and lexical-level representations are obtained through 3 layers of GCN aggregation. ② Mention-level graph: Nodes represent four event mentions, constructing mention-mention edges (no edge exists if there are no co-references) and event-event edges (e.g., mention 1 - mention 2). Lexical-level representations and global semantic representations are fused through a lexical mapping matrix to obtain initial features, which are then aggregated through 2 layers of GCN to obtain mention-level representations. ③ Event-level graph: Nodes represent events A, B, and C, constructing central event association edges (event B has the highest mention frequency and is the central event, connecting A and C) and co-occurrence edges (A and B, B and C). Mention-level representations and global semantic representations are fused to obtain initial features, which are then aggregated through 2 layers of GCN to obtain event-level representations.

[0155] Step 4: Causal Relationship Prediction

[0156] Construct a fusion embedding for events A, B, and C, and build joint representations for event pairs (A,B), (B,C), and (A,C) respectively. Input these representations into an MLP classifier. The final prediction results are: there is a causal relationship between event A and event B (heavy rainfall caused flooding), there is a causal relationship between event B and event C (flooding led to the activation of an emergency drainage plan), and there is no direct causal relationship between event A and event C.

[0157] Step 5: Model Training and Optimization

[0158] The above samples are added to the training set and trained together with other labeled samples. The parameters are optimized by minimizing the cross-entropy loss function, and L2 regularization and Dropout are introduced to prevent overfitting and ensure the model's prediction accuracy.

[0159] Figure 8The causal relationship diagram generated in this example of the present invention uses events such as "heavy rainfall", "waterlogging", and "emergency drainage plan" as nodes. The corresponding causal relationship visualization diagram is generated from the news text "Due to heavy rainfall, waterlogging occurred in many places. The municipal department immediately activated the emergency drainage plan, and the waterlogging situation gradually eased." This forms a causal chain of "event A → event B → event C".

[0160] IV. Verification of Technical Effects

[0161] The event causal relationship recognition method proposed in this invention, based on dual-channel and multi-layer graph convolutional networks, has achieved significant technical progress in document-level event causal relationship recognition tasks by introducing a multi-layer graph convolutional network structure and a dual-channel feature encoding mechanism.

[0162] To verify the technical effectiveness of this invention, the method demonstrated superior performance on publicly available datasets. Comparative experiments were conducted on the EventStoryLine v1.5 dataset, which contains 258 documents, covering 5,334 events and 7,805 event pairs, of which 1,770 event pairs were labeled as having causal relationships. During evaluation, documents on the last two topics were designated as the development set, and the remaining documents were evaluated using a 5-fold cross-validation strategy. As shown in Table 2, the proposed method (DMGCN) achieved an overall F1 score of 60.8%, significantly outperforming other listed baseline methods, especially in the more challenging task of identifying cross-sentence causal relationships. This fully demonstrates the effectiveness of the proposed method in handling document-level long-distance, implicit causal inference problems.

[0163] Table 2: Performance comparison of each model on the EventStoryLine v1.5 dataset

[0164]

[0165] This method constructs graph structures at three levels: lexical, mention, and event, and applies graph convolutional networks to each level. This enables the modeling of complete information features, from local syntactic dependencies to global event interactions, effectively addressing the problem that traditional methods cannot fully capture deep interactions between nodes at the same level. Furthermore, the dual-channel encoding module, through parallel fusion of local contextual features and global document semantics, ensures that the model does not lose key low-dimensional semantic features during information aggregation, providing more robust node representations for subsequent graph reasoning and thus achieving superior event causal relationship recognition.

[0166] In addition, the present invention also includes a news event causal relationship identification system. Figure 6The system's modular architecture diagram shows that the core modules include a text input module, a dual-channel encoding module, a multi-layer graph convolutional network module, and a prediction module, while the extended module is a visualization and interaction module. The system supports end-to-end analysis from text input to causal graph visualization. This application can read long news texts, automatically extract key events, construct causal networks between them, and ultimately present them to the user in an interactive graph structure. This significantly improves the practical value of event causal relationship identification technology in real-world scenarios such as public opinion analysis and emergency management.

[0167] The above embodiments are merely preferred embodiments of the present invention and should not be considered as limiting the scope of the present invention. All equivalent variations and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A method for identifying event causal relationships based on dual-channel and multi-layer graph convolutional networks, characterized in that... Includes the following steps: 1) Text input and preprocessing: Obtain the text of the document to be processed, perform word segmentation, part-of-speech tagging, syntactic analysis, and event extraction based on trigger word recognition and argument extraction to obtain word sequence, syntactic dependency, event mention set and event set; 2) Dual-channel encoding: Based on the word sequence obtained in step 1), construct a word input vector that integrates word embedding, event embedding and mention embedding, and input it in parallel into the BiLSTM encoder and the Transformer encoder to extract the local context feature sequence and the global semantic representation sequence, respectively; 3) Multi-layer graph convolutional network processing: Based on the local context feature sequence and global semantic representation sequence output in step 2), word-level, mention-level and event-level graph structures are constructed sequentially. The feature sequence is used as the initial node features of each graph structure. The information of each graph structure is aggregated step by step through the multi-layer graph convolutional network GCN with residual connections to output event-level node representations. 4) Causal relationship prediction: The event-level node representation output from step 3), the global semantic representation output from step 2), and the word-level representation after mapping transformation are fused to construct an event fusion embedding; the fusion embedding of any event pair is concatenated and input into a multilayer perceptron (MLP) classifier to predict whether there is a causal relationship between the event pairs; 5) Model training and optimization: The above model is trained based on the labeled dataset. The model parameters are optimized by minimizing the cross-entropy loss function, and a regularization mechanism is introduced to prevent overfitting.

2. The event causal relationship recognition method based on dual-channel and multi-layer graph convolutional networks according to claim 1, characterized in that... In step 1), the event extraction adopts a method based on trigger word identification and argument extraction. Trigger word identification uses the pre-trained language model BERT-base to classify and determine verbs and nouns in the text, and selects trigger words with event triggering capabilities. Based on syntactic dependencies, the core arguments corresponding to each trigger word are determined, including agent, patient, time, and place. A complete event mention is composed of "trigger word + core argument". Core reference resolution is performed on all event mentions, and multiple mentions pointing to the same real event are grouped into one category to form an event set.

3. The event causal relationship recognition method based on dual-channel and multi-layer graph convolutional networks according to claim 1, characterized in that... In step 2), the specific steps of the dual-channel encoding are as follows: 2.1) Construct an input vector that fuses information from multiple sources. For the first element in the document... Each word element has an input representation. It is composed of word embeddings, event embeddings, and mention embeddings, and the calculation formula is as follows: in, , and These represent the numbers for words, events, and mentions, respectively. , and This corresponds to the embedding layer; for word embeddings The pre-trained language model BERT and static word vectors GloVe are used for initialization to introduce prior lexical semantic knowledge; the entire document's word input vector sequence... They are then fed into two independent encoding channels in parallel for processing; 2.2) The first channel uses a bidirectional long short-term memory network (BiLSTM) to capture the local contextual semantic features of lexical units; for each lexical position... BiLSTM calculates its forward hidden state respectively. and backward hidden state The computation process follows the recursive update mechanism of LSTM: The output of this channel is the first Local context representation of each word It is formed by splicing together its forward and backward hidden states: The second channel employs a Transformer encoder, utilizing a self-attention mechanism to model the global contextual semantic features of tokens; the self-attention layer calculates the attention weights between token pairs based on the query, key, and value matrix; for the input sequence Its self-attention output The calculation is as follows: in, It is a parameter matrix. It is the hidden layer dimension. It is the scaling factor; 2.3) Self-attention output A feedforward neural network (FFN) is used, which includes two linear transformation layers and a non-linear activation function ReLU to enhance representation capabilities. in, Represents the weight matrix. Represents the bias vector; The global context representation of the output of this channel; 2.4) For each token, the dual-channel encoding module outputs two complementary representation sequences: a local context representation sequence generated by the BiLSTM channels. and the global semantic representation sequence generated by the Transformer channel. These two sequences will serve as the initial features of nodes in different levels of graph structures in subsequent multi-layer graph convolutional network modules, laying the foundation for hierarchical information aggregation in the model.

4. The event causal relationship recognition method based on dual-channel and multi-layer graph convolutional networks according to claim 1, characterized in that... In step 3), the specific steps of the multi-layer graph convolutional network processing include: 3.1) In a lexical graph, nodes are all the terms in a document; seven types of edges are defined to construct the graph topology, including: syntactic dependency edges, reference-intra-reference edges, reference-between edges, event edges, sentence-intra-reference edges, sentence-between edges, and self-looping edges; Lexical-level graphs are sequences of local context representations output by the previous module. As the initial features of the nodes, a multi-layer graph convolutional network is then used to learn the graph; for the th The update formula for nodes in a layered graph convolutional network is as follows: in, It is the adjacency matrix of the word-level graph. and It is the first Layer parameters, It's an activation function; here, a residual connection is introduced, concatenating the outputs of the first two layers as a supplement to the input of the current layer. To mitigate the oversmoothing problem that may occur in deep GCNs; the output of the word-level graph convolutional network is the word representation after deep aggregation. ; 3.2) In the mention-level graph, nodes are event mentions; there are three main types of edges: mention-mention edges, event-event edges, and self-looping edges; these are mapped through a lexical matrix. lexical representation and global semantic representation Perform transformation and fusion to initialize the representation of each event mention node: Initial node representation The information is fed into a mention-level graph convolutional network for propagation and aggregation. Its inter-layer update method is similar to that of a word-level graph convolutional network, using an adjacency matrix. This allows us to aggregate information mentioned by neighbors, thereby obtaining a mention representation that contains document-level information. ; 3.3) In an event-level graph, nodes represent events within a document. First, the central event in the document is determined based on its frequency of mention, and it is connected to all other event nodes. Furthermore, co-occurrence edges are established between event pairs mentioned in the same sentence. The initial representation of event nodes... By mentioning the mapping matrix Reference level and Obtained through fusion: This initial representation It is then fed into an event-level graph convolutional network, based on the adjacency matrix of the event-level graph. The convolution operation is performed, and the final output is an event node representation that reflects the potential causal relationship between events and the topic association. ; By sequentially processing and transmitting information through three hierarchical graph structures, a layer-by-layer abstraction and reasoning process is achieved, from lexical and grammatical information to document themes.

5. The event causal relationship recognition method based on dual-channel and multi-layer graph convolutional networks according to claim 1, characterized in that... In step 4), the specific steps for predicting the causal relationship are as follows: 4.1) Representing by piecing together event-level diagrams Global context representation and word-level information after mapping transformation The feature vector for each event Construct a fused embedded representation: in, yes The corresponding representation, yes The corresponding representation, yes The corresponding representation ensures that the event representation simultaneously contains local details, global context, and structured relationships; 4.2) For any pair of events that needs to be determined, the module will embed the final representation of the two events. and Perform vector concatenation to form a joint representation of the event pair. : in, This represents a vector concatenation operation; It was then fed into a multilayer perceptron for causal classification; 4.3) The MLP performs a series of nonlinear transformations to map the high-dimensional features of the input to the probability space of causal categories, and outputs the probability distribution of each causal category through the Softmax activation function: in, and These are the weight matrix and bias vector of the last layer in an MLP classifier. It is a two-dimensional probability vector, representing the predicted probabilities of "a causal relationship exists" and "a causal relationship does not exist" between event pairs, respectively; 4.4) Model training is performed at the document level; for a single document, the loss function is defined as... The sum of the cross-entropy losses of each event pair is used, and an L2 regularization term is introduced to control model complexity and prevent overfitting. in, Representing the The true causal relationship label for each event pair. It is the probability distribution predicted by the model. It is the regularization coefficient. This represents all trainable parameters of the model.

6. An event causal relationship recognition system based on dual-channel and multi-layer graph convolutional networks, characterized in that... It consists of a text input module, a dual-channel encoding module, a multi-layer graph convolutional network module, and a prediction module, in that order. The text input module is used to receive the text of the document to be processed, perform word segmentation and part-of-speech tagging preprocessing, and output a word sequence. The dual-channel encoding module is used to construct multi-source fused word input vectors, and extracts local contextual features and global semantic representations in parallel using BiLSTM and Transformer. The multi-layer graph convolutional network module is used to construct a three-layer graph structure at the word level, mention level, and event level. It aggregates information from each layer through a multi-layer GCN and outputs an event-level representation. The prediction module is used to fuse multi-source features to construct event embeddings and output causal relationship prediction results through an MLP classifier.

7. The event causal relationship recognition system based on dual-channel and multi-layer graph convolutional networks according to claim 6, characterized in that... It also includes a visualization and interaction module, which is built using general web development technology and the Streamlit framework and is connected to the prediction module. The core functions include: ① Text input interface: supports batch / single text input; ② Result display: displays the extracted events and identified causal pairs in tabular form; ③ Interactive causal graph: the node size and color depth map the frequency of causal relationships of events, supports zooming, dragging, and viewing node details, and edges represent causal relationships.

Images