An event causal identification method, device and equipment based on an iterative graph neural network and a medium

Abstract

The application discloses an event causal relationship identification method and device based on an iterative graph neural network, equipment and a medium, relates to the technical field of artificial intelligence and machine learning, and comprises: performing sentence encoding and event extraction on input text to obtain sentence embedding and event mention results. Then, a multi-granularity context perception mechanism is used to generate sentence-level embedding and document-level embedding of the events. Then, an initial event causal graph structure is constructed and encoded to obtain graph embedding. Finally, through an iterative graph optimization mechanism, the sentence-level embedding, the document-level embedding and the graph embedding are combined to dynamically update the event causal graph structure, so that the accurate identification of the event causal relationship is realized. Through the multi-granularity context perception mechanism and the iterative graph optimization mechanism, the local and global context information is effectively integrated, the accuracy and robustness of the document-level event causal relationship identification are improved, especially when long texts and cross-sentence causal relationships are processed, the application can better adapt to complex document structures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and machine learning, in particular to an event causality recognition method and device based on an iterative graph neural network, equipment and a medium. BACKGROUND

[0002] With the rapid development of text data in the direction of long documents, cross-domain and multi-language, event causality often spans multiple sentences or even the entire document (i.e. document-level event causality), and its recognition needs are increasingly urgent in practical applications such as machine reading comprehension, knowledge graph construction, intelligent question answering systems, etc. The current mainstream technical path for document-level event causality recognition mainly includes three categories: first, based on manually designed features such as lexical signals, temporal features, and co-occurrence relationships, the causality is recognized by rule matching or traditional machine learning models. Second, relying on large language models (such as BERT, RoBERTa) to encode the semantics of the text, generate embedded representations of events and context, and then predict the causality through a classification model. Third, event graph reasoning and prompt fine-tuning methods, event graph reasoning constructs a graph structure with events as nodes and causality as edges, and uses graph embedding models (such as GCN, GAT) for causality reasoning; prompt fine-tuning converts the task into a question-answering or mask prediction task, and guides the pre-training model to learn the causality semantics through a prompt template. In addition, some studies attempt to use large language models to realize few-shot or zero-shot event causality recognition. However, existing deep semantic encoding methods mostly encode the document uniformly, ignoring the unique context environment of different events, resulting in the mixing of redundant noise and ambiguous information in the event representation, making it difficult to capture semantic and relationship contexts with discriminative power, and affecting the accuracy of causality recognition. While event graph reasoning methods can capture global structural associations, they are insufficient in mining deep semantic information; prompt fine-tuning methods are good at modeling semantic associations, but are easily disturbed by long text spans and irrelevant events, and cannot achieve deep fusion of semantic information and structural relationships, making it difficult to detect subtle causality clues and increasing the risk of false correlation misjudgment.

[0003] Therefore, there is an urgent need for a more efficient and accurate recognition method to realize document-level event causality recognition. SUMMARY

[0004] The main purpose of the present application is to provide an event causality recognition method and device based on an iterative graph neural network, and equipment and a medium, aiming to solve the technical problem of how to improve the efficiency and accuracy of document-level event causality recognition.

[0005] To achieve the above-mentioned purpose, the present application provides an event causality recognition method based on an iterative graph neural network, comprising:

[0006] The input text is sentence encoded and event extracted to obtain a sentence embedding result and an event mention result;

[0007] The sentence embedding result and the event mention result are input into a preset event causal recognition model to obtain an event causal relationship recognition result, wherein the preset event causal recognition model comprises a multi-granularity context perception encoding module, an event causal graph encoding module, and an iterative graph optimization module;

[0008] The step of inputting the sentence embedding result and the event mention result into the preset event causal recognition model to obtain the event causal relationship recognition result comprises:

[0009] The sentence embedding result and the event mention result are encoded by the multi-granularity context perception encoding module to obtain a sentence-level embedding result and a document-level embedding result of the event, wherein the multi-granularity context perception encoding module adopts a double-level encoding mode of intra-sentence self-attention aggregation and document-level cross-sentence attention weighting to capture local sentence context and global document context of the event respectively;

[0010] An initial event causal graph structure is constructed;

[0011] The initial event causal graph structure is encoded by the event causal graph encoding module to obtain a graph embedding result;

[0012] The sentence-level embedding result, the document-level embedding result, and the graph embedding result are processed by the iterative graph optimization module to obtain the event causal relationship recognition result.

[0013] In an embodiment, the step of encoding the input text and extracting events to obtain a sentence embedding result and an event mention result comprises:

[0014] The input text is processed by word segmentation to divide into a plurality of text segments;

[0015] The text segments are segmented based on punctuation marks and semantic pause rules to obtain a plurality of independent sentences;

[0016] In each of the independent sentences, a text segment comprising an action, a state change, or an event trigger feature is identified and marked as a potential event;

[0017] Event markers are inserted on both sides of the text segment corresponding to each potential event to obtain a marked sentence;

[0018] Each of the marked sentences is input into an encoder of a preset language model to perform semantic encoding on the tokens of each sentence to obtain vector representations of the tokens in each sentence, and the vector representations are integrated to form a sentence embedding result;

[0019] extracting text content between the start event marker and the end event marker in each of the marked sentences as an event mention result, and recording a position index and a character length of each of the event mention result in a corresponding sentence.

[0020] In an embodiment, the step of encoding the sentence embedding result and the event mention result by the multi-granularity context-aware encoding module to obtain a sentence-level embedding result and a document-level embedding result of an event comprises:

[0021] extracting a token vector corresponding to each of the event mention result by the multi-granularity context-aware encoding module;

[0022] calculating semantic association weights between an event and each token in a corresponding sentence by a self-attention mechanism, taking the token vector as a query vector, and taking the sentence embedding result of the corresponding sentence as a key vector and a value vector;

[0023] performing weighted aggregation on the sentence embedding result of the corresponding sentence based on the semantic association weights to obtain an event-specific sentence semantic vector;

[0024] splicing and mapping the event-specific sentence semantic vector and the token vector of the event mention result to obtain a sentence-level embedding result of the event;

[0025] inputting all the sentence-level embedding results into a bidirectional long short-term memory network to obtain a global semantic hidden state of each sentence and take it as a global sentence embedding;

[0026] constructing a cross-sentence attention mask matrix;

[0027] calculating effective attention weights between each sentence and other sentences by the cross-sentence attention mask matrix;

[0028] performing weighted summation on all the global sentence embeddings based on the effective attention weights to obtain a document-level context vector of the sentence;

[0029] splicing and converting the global sentence embedding and the document-level context vector to obtain a document-level embedding result of the event.

[0030] In an embodiment, the step of constructing an initial event causal graph structure comprises:

[0031] performing semantic parsing on each of the sentences to generate a semantic graph, wherein the semantic graph comprises nodes, edges and edge types, and the nodes comprise event nodes and non-event nodes;

[0032] filtering out nodes marked as event types from the semantic graph as event nodes, and extracting semantic description information and associated edge information of each of the event nodes;

[0033] filtering the connection paths between the event nodes according to a preset effective edge type set to obtain candidate paths, wherein the preset effective edge type set includes edge types corresponding to a means class, a modifier class, an operator class, a preposition class, a core role class, and a situational role class;

[0034] adding a directed edge between corresponding event nodes when a length of the candidate path does not exceed a first preset hop threshold or when the length of the candidate path does not exceed a second preset hop threshold and the path includes an explicit causal clue word or a causal related edge type, the causal related edge type including a reason, a causal relationship, a condition, a purpose, a result, and an effect, and the directed edge including a unidirectional directed edge or a bidirectional directed edge;

[0035] constructing a sentence-level event causal subgraph based on the directed edge and the corresponding event nodes;

[0036] merging all the sentence-level event causal subgraphs to generate an initial event causal graph structure.

[0037] In an embodiment, the step of encoding the initial event causal graph structure by the event causal graph encoding module to obtain a graph embedding result includes:

[0038] reading node and edge information of the initial event causal graph structure by the event causal graph encoding module, generating an adjacency matrix, and determining a neighbor node set of each event node;

[0039] generating an initial embedding vector of each event node according to a preset dimension;

[0040] initializing a number of attention heads, a learnable weight matrix, and an activation function of a graph attention network;

[0041] inputting the initial embedding vector and the adjacency matrix into the graph attention network for calculation to obtain an original attention coefficient between each event node and a neighbor node;

[0042] normalizing the original attention coefficient by a softmax function to obtain a normalized attention weight;

[0043] weighting and aggregating initial embedding vectors of neighbor nodes based on the normalized attention weight to obtain a node aggregated feature vector;

[0044] splicing and linearly transforming the node aggregated feature vectors of all attention heads to obtain a graph embedding result of each event node.

[0045] In an embodiment, the step of processing the sentence-level embedding result, the document-level embedding result and the graph embedding result by the iterative graph optimization module to obtain an event causal relationship recognition result comprises:

[0046] performing weighted summation on the sentence-level embedding result, the document-level embedding result and the graph embedding result of each event by the iterative graph optimization module to obtain an initial event enhanced representation result;

[0047] concatenating the initial event enhanced representation results of any two events into an event pair joint representation;

[0048] inputting the event pair joint representation into a multi-layer perception machine model to output a forward causal probability, a reverse causal probability and a no causal probability of a corresponding event pair;

[0049] comparing the forward causal probability, the reverse causal probability and the no causal probability with a preset probability threshold to determine an effective causal relationship type;

[0050] inputting the effective causal relationship type into an initial event causal graph structure for updating to obtain an updated event causal graph structure;

[0051] inputting the updated event causal graph structure into a graph attention network for re-encoding to obtain an updated graph embedding result;

[0052] unifying dimensions of the sentence-level embedding result, the document-level embedding result and the updated graph embedding result through linear transformation to obtain an updated event representation result;

[0053] outputting a target event causal graph structure until the updated event representation result is less than a preset error threshold;

[0054] based on the target event causal graph structure, extracting causal relationship types of all event pairs to output an event causal relationship recognition result.

[0055] In an embodiment, the step of inputting the event pair joint representation into a multi-layer perception machine model to output a forward causal probability, a reverse causal probability and a no causal probability of a corresponding event pair comprises:

[0056] inputting the event pair joint representation into a first fully connected layer through linear transformation to obtain an intermediate feature vector;

[0057] applying a non-linear activation function to the intermediate feature vector to obtain an activated feature vector;

[0058] inputting the activated feature vector into a second fully connected layer to output three types of relationship original scores, the three types of relationships including a positive causal relationship, a reverse causal relationship and a no causal relationship;

[0059] Applying a softmax function to the three types of relationship raw scores obtains the forward causal probability, the reverse causal probability and the acausal probability of the corresponding event pair.

[0060] In addition, to achieve the above object, the present application also proposes an event causal recognition device based on an iterative graph neural network, which comprises:

[0061] A text processing module is configured to perform sentence encoding and event extraction on input text to obtain a sentence embedding result and an event mention result.

[0062] A recognition module is configured to input the sentence embedding result and the event mention result into a preset event causal recognition model to obtain an event causal relationship recognition result, wherein the preset event causal recognition model comprises a multi-granularity context perception encoding module, an event causal graph encoding module and an iterative graph optimization module; the recognition module is further configured to perform encoding processing on the sentence embedding result and the event mention result by using the multi-granularity context perception encoding module to obtain a sentence-level embedding result and a document-level embedding result of an event, wherein the multi-granularity context perception encoding module adopts a double-level encoding mode of intra-sentence self-attention aggregation and document-level cross-sentence attention weighting to capture local sentence context and global document context of the event respectively; an initial event causal graph structure is constructed; the event causal graph encoding module is configured to perform encoding processing on the initial event causal graph structure to obtain a graph embedding result; and the iterative graph optimization module is configured to process the sentence-level embedding result, the document-level embedding result and the graph embedding result to obtain the event causal relationship recognition result.

[0063] In addition, to achieve the above object, the present application also proposes a storage medium, which is a computer readable medium, and a computer program is stored on the storage medium, wherein the computer program is executed by a processor to implement the steps of the event causal recognition method based on the iterative graph neural network as described above.

[0064] In addition, to achieve the above object, the present application also provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the steps of the event causal recognition method based on the iterative graph neural network as described above.

[0065] The application obtains sentence embedding and event mention results by performing sentence encoding and event extraction on input text. Then, the sentence-level embedding and document-level embedding of the event are generated by using a multi-granularity context perception mechanism. Then, an initial event causal graph structure is constructed and encoded to obtain a graph embedding. Finally, the event causal graph structure is dynamically updated to realize accurate identification of the event causal relationship by combining the sentence-level embedding, the document-level embedding and the graph embedding through an iterative graph optimization mechanism. Through the multi-granularity context perception mechanism and the iterative graph optimization mechanism, the local and global context information is effectively integrated, and the accuracy and robustness of the document-level event causal relationship identification are improved, especially when processing long text and cross-sentence causal relationship, which can better adapt to complex document structures. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor.

[0067] Figure 1 The flowchart of the first embodiment of the event causal recognition method based on the iterative graph neural network of the present application is shown.

[0068] Figure 2 The preset event causal recognition model structure block diagram of the first embodiment of the event causal recognition method based on the iterative graph neural network of the present application is shown.

[0069] Figure 3 The flowchart of the second embodiment of the event causal recognition method based on the iterative graph neural network of the present application is shown.

[0070] Figure 4 The module structure diagram of the event causal recognition device based on the iterative graph neural network of the first embodiment of the event causal recognition method based on the iterative graph neural network of the present application is shown.

[0071] Figure 5 The device structure diagram of the hardware running environment involved in the event causal recognition method based on the iterative graph neural network in the embodiments of the present application is shown.

[0072] The purpose of the present application, the functional characteristics and the advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0073] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application, and are not used to limit the present application.

[0074] For better understanding of the technical solutions of the present application, the following will be described in detail in conjunction with the drawings of the specification and specific embodiments.

[0075] With the rapid development of text data towards long documents, cross-domain, and multi-language, event causal relationships often span multiple sentences or even entire documents (i.e., document-level event causal relationships), and the demand for their identification is increasingly urgent in practical applications such as machine reading comprehension, knowledge graph construction, and intelligent question answering systems. The current mainstream technical path for document-level event causal relationship identification mainly includes three categories: first, based on manually designed features such as lexical signals, temporal features, and co-occurrence relationships, the causal relationship is identified through rule matching or traditional machine learning models. Second, relying on large language models (such as BERT, RoBERTa) to encode the semantics of the text, generating embedded representations of events and context, and then predicting causal relationships through classification models. Third, event graph reasoning and prompt fine-tuning methods, event graph reasoning constructs a graph structure with events as nodes and causal relationships as edges, and uses graph embedding models (such as GCN, GAT) for causal reasoning; prompt fine-tuning converts the task into a question-answering or mask prediction task, guiding the pre-training model to learn causal semantics through a prompt template. In addition, some studies attempt to use large language models to achieve few-shot or zero-shot event causal relationship identification. However, existing deep semantic encoding methods often encode the entire document uniformly, ignoring the unique context of different events, resulting in the mixing of redundant noise and ambiguous information in the event representation, making it difficult to capture discriminative semantic and relationship contexts, and affecting the accuracy of causal relationship identification. While event graph reasoning methods can capture global structural associations, they do not adequately exploit deep semantic information; prompt fine-tuning methods are good at modeling semantic associations, but are easily disturbed by long text spans and unrelated events, and cannot achieve deep integration of semantic information and structural relationships, making it difficult to detect subtle causal clues and increasing the risk of false correlation errors.

[0076] Therefore, the present application proposes a more efficient and accurate identification method to realize document-level event causal relationship identification. The main solution of the embodiments of the present application is: performing sentence encoding and event extraction on the input text to obtain sentence embedding results and event mention results; inputting the sentence embedding results and the event mention results into a preset event causal identification model to obtain event causal relationship identification results, wherein the preset event causal identification model includes a multi-granularity context perception encoding module, an event causal graph encoding module, and an iterative graph optimization module.

[0077] Based on the above, the embodiments of the present application also provide an event causal identification method based on an iterative graph neural network, which is described with reference to Figure 1 , Figure 1 The flowchart of the first embodiment of the event causal identification method based on the iterative graph neural network of the present application is shown in the figure.

[0078] In this embodiment, the event causal recognition method based on the iterative graph neural network comprises steps S10-S50.

[0079] Step S10, sentence encoding and event extraction are performed on the input text to obtain sentence embedding results and event mention results.

[0080] It should be noted that the input text refers to the original text data that needs to be recognized for event causal relationship, and these texts can come from news reports, historical documents, academic papers and other sources, and their content contains descriptions of various events, providing basic data support for obtaining sentence embedding results and event mention results. Sentence encoding is the process of converting each sentence in the input text into a vector that can represent its semantic information, i.e., sentence embedding results. This process is usually implemented with the help of pre-trained language models (such as BERT), which can capture lexical, grammatical and semantic information in the text, providing a basis for subsequent semantic analysis. Event extraction is to identify text segments with action, state change or trigger features from sentences and mark them as event mentions, while recording the specific location and length information of these event mentions in the sentence to obtain event mention results. This process is crucial for understanding events and their relationships in the text.

[0081] Further, step S10 further comprises: first, performing word segmentation processing on the input text to divide a plurality of text segments. Specifically, word segmentation processing is a basic step in natural language processing, and the purpose is to divide continuous text strings into units with independent meaning, i.e., text segments. These text segments can be words, phrases or character sequences, depending on the word segmentation algorithm used and language characteristics. For example, in Chinese text processing, word segmentation processing will divide the sentence "he went to city A" into three text segments "he / went / to city A". Word segmentation processing provides the basic unit for subsequent sentence segmentation and event extraction, enabling the model to process semantic information in the text more finely.

[0082] Next, the text segments are segmented based on punctuation marks and semantic pause rules to obtain a plurality of independent sentences. Specifically, punctuation marks are symbols used in text to mark the end or separation of a sentence, such as periods, question marks, exclamation marks, etc. Semantic pause rules refer to determining the natural segmentation points of sentences based on the semantic structure and context information of the text. These rules can help process text segments that do not have clear punctuation marks, such as continuous dialogues or ancient texts without punctuation. By combining punctuation marks and semantic pause rules, text segments can be further divided into independent sentence units. For example, for the text segment "he went to city A, I am very happy.", it can be divided into two independent sentences: "he went to city A" and "I am very happy." This process ensures the integrity and independence of each sentence, providing clear boundaries for subsequent event extraction.

[0083] Next, text segments that include action, state change, or event trigger features are identified in each independent sentence and marked as potential events. Specifically, potential events refer to text segments in a sentence that may represent events, which often contain actions, state changes, or trigger features. For example, the verb phrase "went to city A" or the noun phrase "earthquake occurred" can be considered as potential events. The process of identifying potential events usually relies on pre-trained language models or specific feature extractors, which can identify key actions or state changes in a sentence. For example, in the sentence "he went to city A", "went to city A" is identified as a potential event. This process provides a foundation for subsequent event labeling and semantic encoding, ensuring that the model can accurately locate and process event information.

[0084] Then, event markers are inserted on both sides of the text segment corresponding to each potential event to obtain the marked sentence. Specifically, event markers are special markers inserted on both sides of the text segment after identifying potential events, which are used to clearly identify the location and range of events. For example, in the sentence "he went to city A", after identifying "went to city A" as a potential event, markers can be inserted on both sides to obtain the marked sentence "he [EVENT] went to city A [ / EVENT]". These markers not only help the model identify the location of the event, but also provide clear boundaries for subsequent semantic encoding. This process enables the model to more accurately process event information, avoiding confusion or loss of event information.

[0085] Each marked sentence is then input into the encoder of the pre-set language model to perform semantic encoding on the tokens of each sentence, obtaining vector representations of each token in the sentence, and integrating to form a sentence embedding result. Specifically, the pre-set language model is a pre-trained model, usually trained based on a large amount of text data, which can perform semantic encoding on tokens (i.e. units after tokenization) in the text to generate vector representations. These vectors can capture the semantic information of the tokens. For example, using the BERT model to encode the sentence "he [EVENT] went to city A [ / EVENT]", the vector representation of each token (such as "he", "went", "city A") can be generated. Then, through some aggregation method (such as taking the average or maximum pooling), these vectors are integrated into a sentence-level vector, i.e. the sentence embedding result. This process provides rich semantic information for subsequent semantic analysis and causal relationship identification, enabling the model to better understand and process the semantic content of the sentence.

[0086] Finally, the text content between the start and end event markers in each tagged sentence is extracted as the event mention result, and the position index and character length of each event mention result in the corresponding sentence are recorded. Specifically, the event mention result is the text content extracted from the tagged sentence between the start and end event markers. For example, in the sentence "He [EVENT] went to City A[ / EVENT]", "went to City A" is extracted as the event mention result. Simultaneously, the position index (such as the start and end positions) and character length (e.g., "went to City A" is 5 characters) of each event mention result in the sentence are recorded. This process not only extracts the specific content of the event but also provides its precise position information in the sentence, facilitating subsequent event analysis and causal relationship identification. This process provides the model with clear event boundaries and positional information, enabling the model to process and analyze event information more accurately.

[0087] By performing word segmentation, sentence segmentation, latent event identification, event tag insertion, semantic encoding, and event mention extraction on the input text, event information can be effectively extracted from the text, generating structured sentence embeddings and event mention results. This process not only provides high-quality input data for subsequent multi-granularity context-aware coding and event causality identification, but also ensures the accuracy and completeness of the event information.

[0088] Step S20: Input the sentence embedding result and the event mention result into the preset event causal recognition model to obtain the event causal relationship recognition result.

[0089] It should be noted that the preset event causality recognition model is a deep learning model specifically designed to identify causal relationships between events in text. It achieves high-precision causal relationship recognition through the collaborative work of multiple modules. For example... Figure 2 The diagram shows the structural block of the preset event causality recognition model. This model includes a multi-granularity context-aware encoding module, an event causality graph encoding module, and an iterative graph optimization module. The multi-granularity context-aware encoding module is responsible for further contextual encoding of sentence embeddings and event mentions, capturing the semantic information of events at the sentence and document levels. The event causality graph encoding module uses these encoding results to construct an event causality graph, representing the potential causal relationships between events in graphical form. The iterative graph optimization module iteratively optimizes the structure and event representation of the event causality graph, gradually improving the accuracy of causal relationship recognition.

[0090] Specifically, after inputting the sentence embedding results and event mention results into the preset event causality recognition model, the input data is first processed by a multi-granularity context perception coding module. The module captures the local sentence context and global document context of the events respectively through intra-sentence self-attention aggregation and document-level cross-sentence attention weighting, and generates high-quality event representations. Then, an event causality graph coding module constructs an initial event causality graph according to the event representations, in which the nodes represent events and the edges represent the potential causal relationships between events. Finally, an iterative graph optimization module iteratively optimizes the event causality graph multiple times, dynamically adjusts the graph structure and event representation, gradually improves the accuracy of causal relationship recognition, and finally outputs the event causality relationship recognition result.

[0091] Further, through analysis of experimental data, the experimental data part describes the experimental verification process of the method of the embodiment in detail. The content of this part covers multiple aspects such as experimental setting, experimental data set, evaluation index, baseline setting, and experimental results and analysis. Three widely used event causality relationship recognition benchmark data sets are selected:

[0092] EventStoryLine v0.9 (ESL) contains 258 news documents covering 22 topics, a total of 5334 event mentions, including 1770 intra-sentence causal event pairs and 3855 cross-sentence causal event pairs. According to the standard annotation specification, event pairs marked as “PRECONDITION” are classified as “cause” relationships, and event pairs marked as “FALLING-ACTION” are classified as “result” relationships. The last two topics are used as the development set according to the standard data division scheme, and the remaining 20 topics are evaluated using 5-fold cross-validation.

[0093] MAVEN-ERE is a large-scale event relationship extraction data set based on diversified Wikipedia topics. The data set uses an enhanced annotation system, covering event coreference, temporal, causal and sub-event relationships, etc. It contains 4480 documents, 103193 events and 57992 causal event pairs. According to the standard annotation convention, “CAUSE” and “PRECONDITION” labels are both considered as causal relationships. The original evaluation protocol is followed, and the original development set is used as the test set.

[0094] MECI is a multilingual event causality identification dataset derived from Wikipedia texts in five languages (English, Danish, Spanish, Turkish, and Urdu). It provides a valuable resource for multilingual identification and cross-lingual transfer learning, containing 3591 documents, 46000 events, and 11000 causal event pairs. Thirty models are selected as baseline methods, covering the following five mainstream technical routes: feature pattern matching-based methods, large language model-based question and answer methods, prompt fine-tuning-based methods, event graph reasoning methods, and multilingual ECI-specific methods.

[0095] According to the experimental results table based on causal directionality identification on the ESL dataset shown in Table 1, the experimental results table based on causal directionality identification on the MAVEN-ERE dataset shown in Table 2, and the experimental results table based on causal directionality identification on the MECI dataset shown in Table 3.

[0096] Table 1 Experimental results table based on causal directionality identification on the ESL dataset

[0097]

[0098] Table 2 Experimental results table based on causal directionality identification on the MAVEN-ERE dataset

[0099]

[0100] Table 3 Experimental results table based on causal directionality identification on the MECI dataset

[0101]

[0102] According to the table content described above, the performance comparison results of MAIGO and the baseline models on the ESL and MAVEN-ERE datasets under two evaluation modes show that in the DECI task (inter-sentence causal reasoning), MAIGO outperforms all baseline models in DI and EI performance on both datasets, with F1 scores improved by 2.3% and 7.6% (ESL dataset) and 5.8% and 7.0% (MAVEN-ERE dataset) compared to the second-best baseline, verifying the superiority of MAIGO in the DECI task. In addition, in terms of overall performance, MAIGO also achieves the most advanced results, with DI and EI F1 scores having an advantage of 0.1% and 7.3% (ESL dataset) and 7.3% and 7.6% (MAVEN-ERE dataset) over the second-best baseline. The experimental results on the MECI dataset show that in the cross-lingual scenario, MAIGO achieves significant performance improvement: it reaches the most advanced level in both languages (0.1% and 5.6% higher than the second-best model), verifying the strong generalization ability of the method in different language environments.

[0103] The embodiment encodes the input text and extracts events to obtain sentence embedding and event mention results. Then, the sentence-level embedding and document-level embedding of the events are generated using a multi-granularity context perception mechanism. Then, an initial event causal graph structure is constructed and encoded to obtain a graph embedding. Finally, the event causal graph structure is dynamically updated to achieve accurate identification of the event causal relationship by iteratively optimizing the graph mechanism combined with the sentence-level embedding, document-level embedding and graph embedding. Through the multi-granularity context perception mechanism and the iterative graph optimization mechanism, the local and global context information is effectively integrated, and the accuracy and robustness of the document-level event causal relationship identification are improved, especially in handling long texts and cross-sentence causal relationships. The method is excellent and can better adapt to complex document structures.

[0104] Based on the first embodiment of the present application, the same or similar contents as the above embodiment one can refer to the above introduction, and the subsequent will not be described. On this basis, please refer to Figure 3 , the event causal relationship identification method based on the iterative graph neural network further comprises steps S201-S204:

[0105] Step S201, the sentence embedding result and the event mention result are encoded and processed by the multi-granularity context perception encoding module to obtain the sentence-level embedding result and the document-level embedding result of the event.

[0106] It should be noted that the multi-granularity context perception encoding module is a double-level encoding mode of intra-sentence self-attention aggregation and document-level cross-sentence attention weighting, which respectively captures the local sentence context and global document context of the event. The intra-sentence self-attention aggregation refers to aggregating the sentence embedding results by using the self-attention mechanism at the sentence level, so as to capture the local context information of the event within the sentence. Specifically, for each event mention, the corresponding sentence embedding result is taken as input, and the correlation weight between the event mention and other words in the sentence is calculated by the self-attention mechanism, and then the sentence-level embedding result of the event is obtained. While the document-level cross-sentence attention weighting refers to aggregating the sentence embedding results by using the cross-sentence attention mechanism at the document level, so as to capture the global context information of the event within the document. Specifically, the sentence embedding results of all sentences are taken as input, and the correlation weight between each sentence is calculated by the cross-sentence attention mechanism, and then the document-level embedding result of the event is obtained.

[0107] Further, step S201 further comprises: first, extracting the token vector corresponding to each event mention result by the multi-granularity context perception encoding module. The token vector refers to converting each word or subword in the event mention result into a vector representation, and these vectors can capture the semantic information of the word. Assuming that the event mention result is , which contains tokens Each token The vector representation is Therefore, the set of token vectors for the event mention results is: The multi-granularity context-aware coding module extracts these token vectors, providing a foundation for subsequent semantic association calculations.

[0108] Next, using the token vector as the query vector and the sentence embedding result of the corresponding sentence as the key and value vectors, a self-attention mechanism is used to calculate the semantic association weights between the event and each token in the corresponding sentence. Specifically, the self-attention mechanism is a method used to calculate the association weights between different parts. Here, the token vector mentioned in the event is used as the query vector, and the sentence embedding result of the corresponding sentence is used as the key and value vectors. Assume the sentence embedding result is... Through the self-attention mechanism, the semantic association weight between the event mention result and other tokens in the sentence embedding result of the corresponding sentence can be calculated. The specific calculation formula is as follows:

[0109]

[0110] in, Indicates the first sentence in the sentence embedding result A vector representation of each token. Indicates the first result mentioned in the event. A vector representation of each token. Indicates the first result mentioned in the event. The token and sentence embedding results of the first The semantic association weight of each token, is the scaling factor used to control the scale of the dot product result, and softmax is the normalization function.

[0111] Next, the sentence embedding results of the corresponding sentences are weighted and aggregated based on semantic association weights to obtain an event-specific sentence semantic vector. Specifically, weighted aggregation refers to the weighted summation of the sentence embedding results according to the calculated semantic association weights. This process generates a vector that represents the semantic information of the event in the sentence, i.e., the event-specific sentence semantic vector, which is specifically represented as follows:

[0112]

[0113] in Represents the semantic vector of a sentence specific to an event. This represents the total number of tokens in the sentence embedding result. Indicates the first sentence in the sentence embedding result A vector representation of each token. Indicates the first result mentioned in the event. The token and sentence embedding results of the first The semantic association weight of each token.

[0114] Subsequently, the semantic vector of the event-specific sentence and the token vector of the event mention results are concatenated and mapped to obtain the sentence-level embedding result of the event. Specifically, concatenation refers to joining two vectors together to form a longer vector. Mapping refers to transforming the concatenated vector into a new vector through a linear transformation or other function to obtain the sentence-level embedding result of the event, specifically expressed as:

[0115]

[0116] in, The sentence-level embedding result representing the event is a vector obtained by concatenation and mapping, which incorporates local sentence semantic information of the event. The weight matrix is ​​a learnable matrix. Indicates the first result mentioned in the event. A vector representation of each token. This represents the semantic vector of a sentence specific to the event. This process generates a sentence-level embedding of the event, which combines the local semantic information of the event mentions with the global semantic information of the sentence. For example, concatenating the semantic vector of the event-specific sentence with the token vector of the event mentions and then performing a linear transformation yields the sentence-level embedding of the event.

[0117] Then, all sentence-level embedding results are input into a Bidirectional Long Short-Term Memory (BiLSTM) network to obtain the global semantic hidden state of each sentence, which is then used as the global sentence embedding. Specifically, a BiLSTM network is a neural network capable of capturing sequential dependencies in sequential data. By inputting all sentence-level embedding results into a BiLSTM, the global semantic hidden state of each sentence can be obtained, specifically represented as follows:

[0118]

[0119] in, This represents the global semantic hidden state of each sentence. This represents the sentence-level embedding results for all events. This represents bidirectional long short-term memory (BiLSTM) processing. These hidden states can capture the semantic relationships between sentences. For example, for a document containing multiple sentences, BiLSTM can capture the semantic information of each sentence in the entire document and generate a global sentence embedding.

[0120] Then construct the cross-sentence attention mask matrix. Cross-sentence attention mask matrix It is a matrix used to control the attention calculation between sentences, with a dimension of . Specifically, it is expressed as:

[0121]

[0122] wherein, the element in the cross-sentence attention mask matrix at the i-th row and the j-th column, taking the value of 1 or 0, is used to identify whether the i-th sentence and the j-th sentence can interact with each other through attention, and whether the i-th sentence and the j-th sentence can interact with each other through attention, represents a graph association identifier, taking the value of 1 to identify the existence of an association, and 0 to identify that the sentences cannot interact with each other.

[0123] It can determine which sentences can pay attention to each other according to the preset rules (such as the distance between sentences or semantic relationship). For example, a mask matrix can be set so that adjacent sentences can pay attention to each other, while sentences that are far apart cannot pay attention to each other.

[0124] Then, the effective attention weights between each sentence and other sentences are calculated through the cross-sentence attention mask matrix. Specifically, the effective attention weights refer to the attention weights between sentences calculated under the constraint of the cross-sentence attention mask matrix, which can reflect the semantic association strength between sentences.

[0125] Based on the effective attention weights, the weighted sum of all global sentence embeddings is performed to obtain the document-level context vector of the sentence. Specifically, the document-level context vector refers to a vector that can represent the context information of the sentence in the entire document. By performing the weighted sum of all global sentence embeddings based on the effective attention weights, the document-level context vector of each sentence can be obtained.

[0126] Finally, the global sentence embedding and the document-level context vector are spliced and converted to obtain the document-level embedding result of the event. Specifically, the document-level embedding result of the event refers to a vector that can represent the semantic information of the event in the entire document. By splicing and converting the global sentence embedding and the document-level context vector, the document-level embedding result of the event can be generated.

[0127] Through this double-level encoding method, the local sentence context and the global document context of the event can be captured at the same time, thereby generating high-quality event representations. This process not only improves the quality of event representations, but also provides rich contextual information for subsequent event causal relationship identification, enabling the model to more accurately understand and analyze the causal relationships between events.

[0128] Step S202, constructing an initial event causal graph structure.

[0129] ​​​It should be noted that the initial event causal graph structure refers to a directed graph constructed based on the events identified in the text and their potential causal relationships. The graph consists of event nodes and causal relationship edges, where the event nodes represent event mentions in the text, and the causal relationship edges represent the causal connections between events. The purpose of constructing the initial event causal graph structure is to provide a basic framework for subsequent causal relationship identification and optimization, so that the model can perform more effective causal reasoning based on the graph structure.

[0130] Further, step S202 includes: first, performing semantic parsing on each sentence to generate a semantic graph, wherein the semantic graph includes nodes, edges and edge types, and the nodes include event nodes and non-event nodes. Specifically, semantic parsing refers to the process of converting natural language sentences into structured semantic representations. The generated semantic graph is a graphical semantic structure, where nodes represent entities, concepts or events in the sentence, edges represent semantic relationships between nodes, and edge types further describe the nature of these relationships. Event nodes refer to nodes representing events, such as "happen", "cause" and the like, while non-event nodes represent other types of entities or concepts, such as "person", "place" and the like. Through semantic parsing, the semantic information of the sentence can be represented in the form of a graph, providing a basis for subsequent event causal relationship identification.

[0131] Next, nodes labeled as event types are selected from the semantic graph as event nodes, and the semantic description information and associated edge information of each event node are extracted. Specifically, nodes labeled as event types in the semantic graph are event nodes. Semantic description information refers to detailed descriptions of event nodes, including event types, participants, etc. Associated edge information refers to edges directly connected to event nodes and their types, which describe the relationship between events and other nodes. By filtering and extracting this information, the specific content of each event node and its semantic role in the sentence can be determined, providing necessary information for subsequent causal relationship analysis.

[0132] Next, the connection paths between the event nodes are filtered according to a preset effective edge type set to obtain candidate paths, wherein the preset effective edge type set includes edge types corresponding to the means class, the modifier class, the operator class, the preposition class, the core role class and the context role class. Specifically, the connection path between the event nodes refers to a route composed of a series of edges and nodes from one event node to another event node in the semantic graph, which is an important form reflecting the indirect semantic association between the event nodes. In this embodiment, the composition of the path is determined by the distribution of the event nodes and the connection of the edges in the semantic graph. In addition, the preset effective edge type set is a set containing edges of specific semantic relationship types, wherein the means class edge type corresponds to an edge representing the means-event relationship, the modifier class corresponds to an edge representing the modifier-modified relationship, the operator class corresponds to an edge representing the operation relationship, the preposition class corresponds to an edge representing the semantic relationship of preposition connection, the core role class corresponds to an edge representing the core role-event relationship, and the context role class corresponds to an edge representing the context role-event relationship. The set provides a judgment standard for path filtering. By filtering according to these preset edge types, candidate paths that may represent the cause-effect relationship can be found, and these paths have the potential to reflect the effective semantic association between the event nodes, thereby providing a basis for subsequent cause-effect relationship judgment.

[0133] Subsequently, when the length of the candidate path does not exceed the first preset hop threshold, or when the length of the candidate path does not exceed the second preset hop threshold and the path includes an explicit causal clue word or a causal related edge type, a directed edge is added between the corresponding event nodes, the causal related edge type includes reason, causal relationship, condition, purpose, result and effect, and the directed edge includes a unidirectional directed edge or a bidirectional directed edge. Specifically, the first preset hop threshold is a hop standard preset for judging whether the length of the candidate path meets the requirement, the hop number refers to the number of edges in the path, and the threshold is used to preliminarily screen out shorter candidate paths with more direct semantic association. In this embodiment, the threshold is set according to actual semantic association analysis requirements to ensure the rationality of the screening. In addition, the second preset hop threshold is another preset hop standard, which usually has a value greater than the first preset hop threshold, and is used to relax the restriction on the path length under certain conditions, so that part of the longer paths containing key causal clues are not missed. The explicit causal clue word refers to a word directly indicating the causal relationship, such as "because" and "therefore". The causal related edge type refers to the edge type related to the causal relationship, including reason, causal relationship, condition, purpose, result and effect. According to these conditions, a directed edge can be added between the event nodes to represent the causal relationship. The direction of the directed edge represents the direction of the causal relationship, which can be unidirectional or bidirectional, depending on the nature of the causal relationship. The length of each candidate path is checked one by one, and if the path length does not exceed the first preset hop threshold, a directed edge is directly added between the corresponding event nodes; if the path length does not exceed the second preset hop threshold and there is an explicit causal clue word or a causal related edge type in the path, a unidirectional or bidirectional directed edge is also added between the corresponding event nodes.

[0134] Then, the sentence-level event causal subgraph is constructed based on the directed edge and the corresponding event node. Specifically, the sentence-level event causal subgraph is constructed based on the event nodes in a single sentence and the added directed edge, and only includes the event nodes with causal relationship in the current sentence and the causal association therebetween, which can accurately reflect the event causal structure in a single sentence. All event nodes with added directed edges in a single sentence are collected, the event nodes are used as subgraph nodes, and the added directed edges are used as subgraph edges, to construct the sentence-level event causal subgraph corresponding to the sentence.

[0135] Then, all the sentence-level event causal subgraphs are merged to generate an initial event causal graph structure. Specifically, the sentence-level event causal subgraphs generated by all sentences are summarized, the event nodes and directed edges in these subgraphs are integrated, the same event nodes and associated edges are merged, and finally an initial event causal graph structure containing all sentence event causal relationships is generated.

[0136] Finally, the sentence semantic structure is structured by semantic parsing and semantic graph construction, the event causal correlation in the sentence is accurately extracted through multiple rounds of screening and causal relationship judgment, the global causal information is integrated through sub-graph merging, and the initial event causal graph structure can comprehensively and accurately reflect the causal relationship between events in the text, thereby providing reliable structured data support for subsequent causal analysis tasks.

[0137] In step S203, the initial event causal graph structure is encoded by an event causal graph encoding module to obtain a graph embedding result.

[0138] It should be noted that the event causal graph encoding module is a module specially used for processing and representing the event causal graph structure, and the purpose is to convert the event nodes and causal relationship edges in the initial event causal graph into a graph embedding result in the form of a vector. The event causal graph encoding module is usually implemented based on a graph neural network (GNN), such as a graph attention network (GAT) or a graph convolutional network (GCN). These networks can update the representation of a node by aggregating the neighborhood information of the node, thereby generating a graph embedding result.

[0139] Further, step S203 includes: first, reading the node and edge information of the initial event causal graph structure by the event causal graph encoding module, generating an adjacency matrix and determining the neighbor node set of each event node. Specifically, the event causal graph encoding module first reads all the event node information and edge information in the initial event causal graph structure, constructs an adjacency matrix according to the edge connection relationship between nodes, i.e., when there is an edge connection between two event nodes, the element at the corresponding position in the adjacency matrix is marked as existing connection, otherwise it is marked as non-existent connection; at the same time, for each event node, all event nodes directly connected to the node through the edge are screened out, and these nodes are integrated to form the neighbor node set of the event node.

[0140] Next, the initial embedding vector of each event node is generated according to the preset dimension. Specifically, the preset dimension is a fixed numerical value determined according to the complexity of the event semantic and the processing capability of the subsequent graph attention network. In addition, the initial embedding vector is an initial numerical vector generated for each event node according to the preset dimension, which is a preliminary quantitative representation of the semantic features of the event node, and can convert the abstract semantic information of the event node into a numerical form that can be processed by a computer, which is usually obtained by a large language model (such as BERT) or random initialization.

[0141] Subsequently, the number of attention heads, learnable weight matrices, and activation functions of the graph attention network are initialized. Specifically, the graph attention network is a graph neural network based on attention mechanism, which is specifically designed for processing graph-structured data. Its main parameters include the number of attention heads, learnable weight matrices, and activation functions. The number of attention heads refers to the number of heads for parallel computing attention in the graph attention network. Multi-head attention mechanism allows the model to learn the relationships between nodes from different perspectives, thereby improving the model's expressive power and flexibility. For example, the number of attention heads can be set to 8 or 16. Each attention head independently computes the attention coefficients between nodes, and then the results of all heads are concatenated or averaged to obtain the final node representation. The learnable weight matrix is a key parameter in the graph attention network, which is used to convert the feature vector of a node into the query, key, and value vectors required by the attention mechanism. In addition, the activation function is used to introduce nonlinearity, allowing the model to learn more complex feature representations. Common activation functions include ReLU, LeakyReLU, etc. In the graph attention network, the activation function is usually applied to the aggregated feature vector of a node to enhance the model's expressive power.

[0142] Subsequently, the initial embedding vectors and adjacency matrices are input into the graph attention network for computation to obtain the original attention coefficients between each event node and its neighbor nodes. Specifically, the initial embedding vectors and adjacency matrices are input into the graph attention network, and the network computes the original attention coefficients between each event node and its neighbor nodes through the attention mechanism. For each event node with its initial embedding vector is converted into a new feature vector through a learnable weight matrix This process can be represented by the following formula:

[0143]

[0144] where is the learnable weight matrix used to convert the initial embedding vector into a feature vector suitable for attention computation. Then for each event node and its neighbor nodes , the original attention coefficient is calculated, which is computed through a shared attention mechanism, usually implemented using a single-layer neural network. The specific formula is as follows:

[0145]

[0146] where is the learnable vector used to compute the attention weight denotes the transposed vector of the learnable vector, denotes the vector concatenation operation, denotes an activation function. These raw attention coefficients reflect the relevance between nodes and are used for subsequent weighted aggregation operations.

[0147] The raw attention coefficients are then normalized by a softmax function to obtain normalized attention weights. Specifically, the softmax function is a commonly used normalization method that converts a set of numerical values into a probability distribution such that the sum of all numerical values is 1. In the graph attention network, the softmax function is used to normalize the raw attention coefficients between each event node and its neighbor nodes, thereby obtaining normalized attention weights. For each event node and its neighbor nodes , the raw attention coefficients are normalized to obtain normalized attention weights , and the specific formula is as follows:

[0148]

[0149] wherein is the neighbor node set of node , exp is the exponential function, denotes the sum of the exponential values of the raw attention coefficients for event node and all neighbor nodes , ensuring that the sum of the attention weights for each node is 1, i.e. . The normalized attention weights reflect the importance of each neighbor node to the current node and are used for subsequent weighted aggregation operations.

[0150] Then, based on the normalized attention weights, the initial embedding vectors of the neighbor nodes are weighted and aggregated to obtain node aggregation feature vectors. Specifically, for each event node , its aggregation feature vector is obtained by weighted aggregation of the initial embedding vectors of its neighbor nodes, and the specific formula is as follows:

[0151]

[0152] wherein is the neighbor node set of node , is the normalized attention weight, is a learnable weight matrix, is an activation function, denotes neighbor node the initial embedding vectors. The aggregated feature vectors are new representations of the nodes that incorporate both the information of the nodes themselves and their neighbor nodes. Through weighted aggregation, the model can dynamically adjust the contribution of each neighbor node, making more important neighbor nodes have a greater impact on the current node. This aggregation method can capture the complex relationships between nodes, generating richer node representations.

[0153] Finally, the node aggregated feature vectors of all attention heads are concatenated and linearly transformed to obtain the graph embedding results of each event node. Specifically, for each event node , assuming there are attention heads, the aggregated feature vector generated by each attention head is , where represents the th attention head. These feature vectors are concatenated to obtain a longer feature vector.

[0154] In order to convert the concatenated feature vector into the final graph embedding result, linear transformation is needed. The graph embedding result is the final representation of the event node, which integrates semantic information extracted from multiple attention heads. This representation not only contains the information of the node itself, but also integrates the information of its neighbor nodes, enabling the model to more comprehensively understand the semantic information of each node. The final graph embedding result can be used for subsequent causal relationship reasoning and other graph-related tasks.

[0155] Step S204, the sentence-level embedding result, the document-level embedding result, and the graph embedding result are processed by the iterative graph optimization module to obtain the event causal relationship recognition result.

[0156] It should be noted that the iterative graph optimization module gradually adjusts the structure of the event causal graph and the event embedding through multiple iterations to improve the accuracy of causal relationship recognition. The iterative optimization process combines the sentence-level embedding result, the document-level embedding result, and the graph embedding result, fully utilizing local, global, and structured information to generate high-quality event representations and causal relationship graphs. The sentence-level embedding result is the vector representation obtained after processing by the multi-granularity context-aware encoding module, which contains the semantic features of the event in the local sentence context. The document-level embedding result is another type of vector representation output by the multi-granularity context-aware encoding module, which carries the semantic information of the event in the global document context. In addition, the graph embedding result is the vector obtained by encoding the initial event causal graph structure by the event causal graph encoding module, which contains the structural association features of the event. Further, the event causal relationship recognition result is the final output of the module, which explicitly defines the causal relationship types between all pairs of events, including forward causal relationship, reverse causal relationship, and no causal relationship.

[0157] Understandably, the iterative graph optimization module first performs a weighted summation of the three types of embedding results to obtain the initial event augmentation representation result. It then concatenates the results of any two events and inputs them into the multilayer perceptron model. The module outputs three types of causal probabilities and filters out valid causal relationship types. Based on this, it updates the initial event causal graph structure and re-encodes it to obtain the updated graph embedding result. The three types of embedding results are then fused to generate the updated event representation result. The above process is repeated until the iteration termination condition is met. Finally, based on the target event causal graph structure, the causal relationship types of all event pairs are extracted to obtain the event causal relationship identification result.

[0158] Further, step S204 includes: First, by using an iterative graph optimization module, a weighted sum is performed on the sentence-level embedding results, document-level embedding results, and graph embedding results for each event to obtain the initial event augmentation representation results. Specifically, for each event... Its sentence-level embedding result is Document-level embedding results The image embedding result is By weighted summing of these embedding results, the initial event augmentation representation is obtained. The specific formula is as follows:

[0159]

[0160] in, These are learnable weights used to adjust the contribution ratio of different embedding results. These parameters are continuously updated during training to optimize model performance. The initial event augmentation representation is a comprehensive representation of the event, fusing semantic information at the sentence, document, and graph structure levels. This multi-granular representation can more comprehensively capture the semantic features of the event, providing richer information for subsequent causal reasoning.

[0161] Next, the initial event augmentation representations of any two events are concatenated to form a joint representation of the event pair. Specifically, the joint representation of the event pair refers to concatenating the initial event augmentation representations of two events to generate a joint vector that can simultaneously represent the semantic information of both events. The specific formula is as follows:

[0162]

[0163] in, Indicates an event The initial event enhancement representation, Indicates an event The initial event enhancement representation, This indicates a splicing operation. Indicates an event and events The joint representation of event pairs is a vector that integrates the semantic information of two events, capturing the semantic association and potential causal relationship between them. Through this joint representation, the model can more effectively predict causal relationships and determine whether a causal relationship exists between two events.

[0164] Next, the joint representation of event pairs is input into the multilayer perceptron model, which outputs the forward causal probability, backward causal probability, and non-causal probability of the corresponding event pairs. It should be noted that a multilayer perceptron (MLP) is a feedforward neural network composed of multiple fully connected layers, used to learn complex nonlinear mapping relationships in input data. In this step, the joint representation of event pairs is input into the MLP model, which learns patterns and relationships in the data and outputs the forward causal probability, backward causal probability, and non-causal probability of the corresponding event pairs. These probability values ​​reflect the direction and strength of the possible causal relationship between the event pairs. Further, the joint representation of event pairs is input into the first fully connected layer and subjected to a linear transformation to obtain an intermediate feature vector. The specific formula is as follows:

[0165]

[0166] in, This is the weight matrix of the first fully connected layer. This is the bias vector.

[0167] By applying a non-linear activation function to the intermediate feature vector, the activated feature vector is obtained. The specific formula is as follows:

[0168]

[0169] Among them, is the activation function. The activated feature vector is a feature representation after nonlinear processing, which can capture the nonlinear relationships in the input features and provide richer information for subsequent causal relationship prediction.

[0170] The activated feature vectors are input into the second fully connected layer, which outputs the original scores of the three classes of relationships. The specific formula is as follows:

[0171]

[0172] Among them, This is the weight matrix of the second fully connected layer. This is the bias vector. The three types of relationships mentioned above include positive causality, negative causality, and no causality; therefore, the original scores for the three types of relationships are... is a three-dimensional vector, representing the original scores of the positive causal relationship, the reverse causal relationship and the non-causal relationship of the event pair. The softmax function is applied to the original scores of the three types of relationships to obtain the forward causal probability, the reverse causal probability and the non-causal probability of the corresponding event pair. The forward causal probability is a probability value representing the existence of a causal relationship in which the former event leads to the latter event. The higher the value, the greater the possibility of the existence of the causal relationship in this direction. In addition, the reverse causal probability is a probability value representing the existence of a causal relationship in which the latter event leads to the former event, which is used to judge the possibility of the causal relationship in the opposite direction of the forward direction. The non-causal probability is a probability value representing the non-existence of any causal relationship between the two events. The higher the value, the greater the possibility of the non-existence of the causal relationship between the two events.

[0173] Subsequently, the forward causal probability, the reverse causal probability and the non-causal probability are compared with a preset probability threshold to determine the effective causal relationship type. It should be noted that the preset probability threshold is a critical value that is preset to judge whether the causal relationship is effective, and is the core standard for distinguishing between effective and ineffective causal relationships. Through the threshold, causal relationships with statistical significance can be screened out. The effective causal relationship type is the real existing causal relationship category determined after probability comparison, including forward causal relationship, reverse causal relationship and non-causal relationship. Specifically, the calculated forward causal probability, reverse causal probability and non-causal probability are compared with the preset probability threshold, respectively. If the forward causal probability is higher than the threshold, it is determined as a forward causal relationship. If the reverse causal probability is higher than the threshold, it is determined as a reverse causal relationship. If none of them is higher than the threshold or the non-causal probability is higher than the threshold, it is determined as a non-causal relationship. Finally, the effective causal relationship type is obtained.

[0174] The effective causal relationship type is then input into the initial event causal graph structure for updating to obtain an updated event causal graph structure. It should be understood that the effective causal relationship type obtained through probability screening is received, and based on the initial event causal graph structure, new directed edges are added according to the effective causal relationship type, and invalid directed edges that do not match the determination result are deleted, to complete the adjustment of the graph structure and obtain the updated event causal graph structure. Specifically, if a new causal relationship (such as forward causal or reverse causal) is identified, the corresponding directed edge is added in the graph. If it is identified that the causal relationship corresponding to an edge does not exist (such as non-causal relationship), the edge is deleted from the graph. If the direction of the causal relationship is inconsistent with the direction of the existing edge in the graph, the direction of the edge is modified.

[0175] Then, the updated event causal graph structure is input into a graph attention network for re-encoding to obtain an updated graph embedding result. Specifically, the updated event causal graph structure is taken as input and passed into the graph attention network. The network first calculates the original attention coefficients of each event node and neighbor nodes and normalizes them, then aggregates the neighbor node features based on the normalized weights, and outputs the vector representation of each event node after linear transformation to obtain the updated graph embedding result.

[0176] The sentence-level embedding result, the document-level embedding result, and the updated graph embedding result are fused by linear transformation to unify the dimensions and obtain an updated event representation result. The fusion of the above-mentioned sentence-level embedding result, the document-level embedding result, and the updated graph embedding result is a process of combining vectors of three different dimensions and different feature types, aiming to integrate the local semantic features, global document features, and the latest structure correlation features of the event to form a more comprehensive event representation.

[0177] The iteration is stopped until the updated event representation result is less than a preset error threshold, and the target event causal graph structure is output. It should be noted that the preset error threshold is a critical value preset for judging whether the iteration is terminated, which is determined based on the model recognition accuracy requirement and the actual application scenario, and is the core standard for measuring the optimization degree of the event representation result. In addition, the updated event representation result is the vector representation obtained by fusing the three types of embedding results and linear transformation in each iteration, and the change amplitude reflects the optimization degree of the event features. Specifically, after obtaining the updated event representation result in each iteration, the difference value between the result and the event representation result of the last iteration is calculated, and the difference value is compared with the preset error threshold. If the difference value is less than the preset error threshold, it means that the event representation result has stabilized and the optimization space is limited, at which point the iteration is stopped and the current event causal graph structure is output as the target event causal graph structure.

[0178] Finally, based on the target event causal graph structure, the causal relationship types of all event pairs are extracted, and an event causal relationship recognition result is output. It should be noted that an event pair refers to a combination of any two event nodes in the target event causal graph structure, and each combination corresponds to a pair of events that may have a causal relationship. The causal relationship type is the real existing relationship between the event pairs, including forward causal relationship, reverse causal relationship, and no causal relationship, which is a clear definition of the causal logic between events. In addition, the event causal relationship recognition result is the final output formed by integrating the causal relationship types of all event pairs, which comprehensively presents the causal relationship between all events in the input text. Specifically, all event nodes in the target event causal graph structure are traversed to generate all possible event pairs, the causal relationship type corresponding to each event pair is determined according to the existence and direction of the directed edge between the event pairs, all causal relationship types of the event pairs are integrated, and the event causal relationship recognition result is output.

[0179] The embodiment obtains the sentence-level and document-level embedding results of the event by the double-level encoding of the multi-granularity context perception encoding module, comprehensively captures the local and global context, reconstructs the initial event causal graph, and generates the graph embedding result through the event causal graph encoding module. Finally, the three types of embedding results are fused through the iterative graph optimization module to dynamically optimize the graph structure and event representation until the termination condition is met, and the event causal relationship recognition result is output, realizing the deep fusion of semantic features and structural features, effectively solving the long-distance and implicit causal relationship recognition problem in long text, and improving the recognition accuracy.

[0180] Based on the first embodiment of the present application, the present application also provides an event causal recognition device based on an iterative graph neural network, please refer to Figure 4 , the device comprises:

[0181] The text processing module 10 is configured to perform sentence encoding and event extraction on the input text to obtain sentence embedding results and event mention results.

[0182] The recognition module 20 is configured to input the sentence embedding results and the event mention results into a preset event causal recognition model to obtain an event causal relationship recognition result, wherein the preset event causal recognition model comprises a multi-granularity context perception encoding module, an event causal graph encoding module, and an iterative graph optimization module. The recognition module 20 is further configured to perform encoding processing on the sentence embedding results and the event mention results through the multi-granularity context perception encoding module to obtain sentence-level embedding results and document-level embedding results of the event. The multi-granularity context perception encoding module adopts a double-level encoding mode of intra-sentence self-attention aggregation and document-level cross-sentence attention weighting to capture the local sentence context and the global document context of the event, respectively. The recognition module 20 is further configured to reconstruct an initial event causal graph structure, perform encoding processing on the initial event causal graph structure through the event causal graph encoding module to obtain a graph embedding result, and perform processing on the sentence-level embedding results, the document-level embedding results, and the graph embedding result through the iterative graph optimization module to obtain the event causal relationship recognition result.

[0183] The event causal recognition device based on the iterative graph neural network provided by the present application adopts the event causal recognition method based on the iterative graph neural network in the above embodiments, which can solve the technical problem of how to improve the efficiency and accuracy of document-level event causal relationship recognition. Compared with the prior art, the event causal recognition device based on the iterative graph neural network provided by the present application has the same beneficial effects as the event causal recognition method based on the iterative graph neural network provided by the above embodiments, and other technical features in the event causal recognition device based on the iterative graph neural network are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0184] The application provides an event causal identification device based on an iterative graph neural network, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the event causal identification method based on the iterative graph neural network in the above-mentioned embodiment one.

[0185] As shown in Figure 5 The event causal identification device based on the iterative graph neural network can include a processing device 1001 (for example, a central processor, a graphics processor, etc.) which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 1002 or programs loaded from a storage device 1003 to a random access memory (RAM) 1004. In the RAM 1004, various programs and data required for the operation of the event causal identification device based on the iterative graph neural network are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the event causal identification device based on the iterative graph neural network to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows the event causal identification device based on the iterative graph neural network with various, it should be understood that all the shown are not required to be implemented or possessed. More or less can be alternatively implemented or possessed.

[0186] In particular, according to the embodiments disclosed in the application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments disclosed in the application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program codes for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from the network through the communication device, or installed from the storage device 1003, or installed from the ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the methods of the embodiments disclosed in the application are performed.

[0187] The event causal relationship identification device based on the iterative graph neural network provided in the application adopts the event causal relationship identification method based on the iterative graph neural network in the above embodiment, and can solve the technical problem of how to improve the efficiency and accuracy of document-level event causal relationship identification. Compared with the prior art, the event causal relationship identification device based on the iterative graph neural network provided in the application has the same beneficial effects as the event causal relationship identification method based on the iterative graph neural network provided in the above embodiment, and other technical features in the event causal relationship identification device based on the iterative graph neural network are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0188] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0189] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0190] The present application provides a computer readable medium having computer readable program instructions (i.e. computer programs) stored thereon, the computer readable program instructions being used to execute the event causal relationship identification method based on the iterative graph neural network in the above embodiment.

[0191] The computer readable medium provided in the present application may, for example, be a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor, or device, or any combination of the above. More specific examples of the computer readable medium can include, but are not limited to, an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer readable medium can be any tangible medium containing or storing a program that can be executed by or used in conjunction with an instruction execution device. The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to an electric wire, an optical cable, an RF (Radio Frequency), and the like, or any suitable combination of the above.

[0192] The above computer readable medium can be contained in the event causal identification device based on the iterative graph neural network; or can exist independently without being assembled into the event causal identification device based on the iterative graph neural network.

[0193] The above computer readable medium carries one or more programs, which, when executed by the event causal identification device based on the iterative graph neural network, enable the event causal identification device based on the iterative graph neural network to be written in one or more programming languages or combinations thereof for executing computer program codes for performing the operations of the present application, the programming languages including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on a user computer, partially on a user computer, as an independent software package, partially on a user computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, through the Internet using an Internet service provider).

[0194] The flowcharts and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of various embodiments of the methods and computer program products according to the present application. In this regard, all block s in the flowcharts and block diagrams represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions ("instructions"). It should also be noted that in some alternative implementations, the functions noted in the block s can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may be executed in the reverse order, depending on the functionality involved. It will also be noted that all the blocks s and combinations thereof can be implemented by hardware based systems that are programmable by a microprocessor or other processing circuitry. As will be realized, the present application is capable of modifications in various obvious aspects, all without departing from the scope of the application. Accordingly, the drawings and descriptions are to be regarded as illustrative in nature, and not as restrictive.

[0195] The modules involved in the embodiments of the present application can be implemented in the form of software or hardware. In some cases, the names of the modules do not limit the units themselves.

[0196] The readable medium provided by the present application is a computer readable medium, which stores computer readable program instructions (i.e., computer programs) for executing the event causal relationship identification method based on the iterative graph neural network. The computer readable program instructions can solve the technical problem of how to improve the efficiency and accuracy of document-level event causal relationship identification. Compared with the prior art, the computer readable medium provided by the present application has the same beneficial effects as the event causal relationship identification method based on the iterative graph neural network provided by the above-mentioned embodiments, which will not be repeated here.

[0197] The present application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of the event causal relationship identification method based on the iterative graph neural network are realized.

[0198] The computer program product provided by the present application can solve the technical problem of how to improve the efficiency and accuracy of document-level event causal relationship identification. Compared with the prior art, the computer program product provided by the present application has the same beneficial effects as the event causal relationship identification method based on the iterative graph neural network provided by the above-mentioned embodiments, which will not be repeated here.

[0199] The above only describes some embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation made by referring to the content of the specification and drawings of the present application, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A method for event causal identification based on iterative graph neural networks, characterized in that, include: Sentence encoding and event extraction are performed on the input text to obtain sentence embedding results and event mention results; The sentence embedding result and the event mention result are input into a preset event causal recognition model to obtain the event causal relationship recognition result. The preset event causal recognition model includes a multi-granularity context-aware encoding module, an event causal graph encoding module, and an iterative graph optimization module. The step of inputting the sentence embedding result and the event mention result into a preset event causal recognition model to obtain the event causal relationship recognition result includes: The sentence embedding result and the event mention result are encoded by the multi-granularity context-aware encoding module to obtain the sentence-level embedding result and document-level embedding result of the event. The multi-granularity context-aware encoding module adopts a two-level encoding method of in-sentence self-attention aggregation and document-level cross-sentence attention weighting to capture the local sentence context and global document context of the event, respectively. Construct the initial event cause-effect graph structure; The initial event cause-effect graph structure is encoded by the event cause-effect graph encoding module to obtain the graph embedding result; The iterative graph optimization module processes the sentence-level embedding results, the document-level embedding results, and the graph embedding results to obtain the event causal relationship identification results.

2. The method as described in claim 1, characterized in that, The steps of performing sentence encoding and event extraction on the input text to obtain sentence embedding results and event mention results include: The input text is segmented into multiple text fragments. The text segment is segmented based on punctuation and semantic pause rules to obtain multiple independent sentences; In each of the independent sentences, identify text fragments that include features such as actions, state changes, or event triggers, and mark them as potential events; Insert event markers on both sides of the text segment corresponding to each potential event to obtain the marked sentence; Each tagged sentence is input into the encoder of a preset language model to perform semantic encoding on the tokens of each sentence, thereby obtaining the vector representation of each token in each sentence, and integrating them to form the sentence embedding result; Extract the text content between the start event marker and the end event marker in the sentence following each marker, as the event mention result, and record the position index and character length of each event mention result in the corresponding sentence.

3. The method as described in claim 1, characterized in that, The step of encoding the sentence embedding result and the event mention result using the multi-granularity context-aware encoding module to obtain the sentence-level embedding result and document-level embedding result of the event includes: The token vector corresponding to each event mention result is extracted using the multi-granularity context-aware coding module. The token vector is used as the query vector, and the sentence embedding result of the corresponding sentence is used as the key vector and value vector. The semantic association weight between the event and each token in the corresponding sentence is calculated through a self-attention mechanism. The sentence embedding results of the corresponding sentences are weighted and aggregated based on the semantic association weights to obtain the event-specific sentence semantic vector; The semantic vector of the event-specific sentence and the token vector of the event mention result are concatenated and mapped to obtain the sentence-level embedding result of the event; All the sentence-level embedding results are input into a bidirectional long short-term memory network to obtain the global semantic hidden state of each sentence, and used as the global sentence embedding. Construct a cross-sentence attention mask matrix; The effective attention weights between each sentence and other sentences are calculated using the cross-sentence attention mask matrix. Based on the effective attention weights, all the global sentence embeddings are weighted and summed to obtain the document-level context vector of the sentence; The global sentence embedding and the document-level context vector are concatenated and transformed to obtain the document-level embedding result of the event.

4. The method as described in claim 1, characterized in that, The steps for constructing the initial event cause-effect graph structure include: Semantic parsing is performed on each sentence to generate a semantic graph, wherein the semantic graph includes nodes, edges and edge types, and the nodes include event nodes and non-event nodes; Nodes marked as event types are selected from the semantic graph as event nodes, and semantic description information and associated edge information of each event node are extracted. The connection paths between the event nodes are filtered according to a preset set of valid edge types to obtain candidate paths. The preset set of valid edge types includes edge types corresponding to means, modifiers, operators, prepositions, core roles, and context roles. When the length of the candidate path does not exceed the first preset hop count threshold, or when the length of the candidate path does not exceed the second preset hop count threshold and the path includes explicit causal clue words or causal related edge types, a directed edge is added between the corresponding event nodes. The causal related edge types include cause, causal relationship, condition, purpose, result and effect. The directed edge includes a one-way directed edge or a two-way directed edge. Construct a sentence-level event causal subgraph based on the directed edges and corresponding event nodes; All the sentence-level event causal subgraphs are merged to generate the initial event causal graph structure.

5. The method as described in claim 1, characterized in that, The step of encoding the initial event cause-effect graph structure through the event cause-effect graph encoding module to obtain the graph embedding result includes: The event causal graph encoding module reads the node and edge information of the initial event causal graph structure, generates an adjacency matrix, and determines the set of neighboring nodes for each event node. Generate an initial embedding vector for each event node according to preset dimensions; Initialize the number of attention heads, learnable weight matrix, and activation function of the graph attention network; The initial embedding vector and the adjacency matrix are input into the graph attention network for calculation to obtain the original attention coefficients between each event node and its neighboring nodes; The original attention coefficients are normalized using the softmax function to obtain normalized attention weights; The initial embedding vectors of neighboring nodes are weighted and aggregated based on the normalized attention weights to obtain the node aggregated feature vector. The node aggregate feature vectors of all attention heads are concatenated and linearly transformed to obtain the graph embedding result for each event node.

6. The method as described in claim 1, characterized in that, The step of processing the sentence-level embedding results, the document-level embedding results, and the graph embedding results through the iterative graph optimization module to obtain the event causality recognition results includes: The iterative graph optimization module performs a weighted summation of the sentence-level embedding result, the document-level embedding result, and the graph embedding result for each event to obtain the initial event augmentation representation result. Concatenate the initial event-enhanced representations of any two events into a joint representation of event pairs; The joint representation of the event pairs is input into the multilayer perceptron model, which outputs the forward causal probability, reverse causal probability, and non-causal probability of the corresponding event pairs. The forward causal probability, reverse causal probability, and no-causal probability are compared with a preset probability threshold to determine the type of valid causal relationship; The effective causal relationship type is input into the initial event causal graph structure for updating, resulting in the updated event causal graph structure; The updated event causal graph structure is input into the graph attention network and re-encoded to obtain the updated graph embedding result; The sentence-level embedding result, the document-level embedding result, and the update graph embedding result are integrated and their dimensions are unified through linear transformation to obtain the update event representation result; The target event cause-effect graph structure is output until the updated event representation result is less than a preset error threshold. Based on the target event causal graph structure, the causal relationship types of all event pairs are extracted, and the event causal relationship identification results are output.

7. The method as described in claim 6, characterized in that, The step of inputting the joint representation of the event pairs into a multilayer perceptron model and outputting the forward causal probability, backward causal probability, and non-causal probability of the corresponding event pairs includes: The joint representation of the event pairs is input into the first fully connected layer and subjected to a linear transformation to obtain an intermediate feature vector; The intermediate feature vector is processed by a nonlinear activation function to obtain the activated feature vector; The activated feature vector is input into the second fully connected layer, which outputs the original scores of three types of relationships, including positive causal relationships, negative causal relationships, and no causal relationships. Applying the softmax function to the original scores of the three types of relationships yields the forward causal probability, reverse causal probability, and non-causal probability of the corresponding event pairs.

8. An event causal recognition device based on an iterative graph neural network, characterized in that, The device includes: The text processing module is used to encode sentences and extract events from the input text, and to obtain sentence embedding results and event mention results. The recognition module is used to input the sentence embedding results and the event mention results into a preset event causal recognition model to obtain the event causal relationship recognition result. The preset event causal recognition model includes a multi-granularity context-aware encoding module, an event causal graph encoding module, and an iterative graph optimization module. It is also used to encode the sentence embedding results and the event mention results through the multi-granularity context-aware encoding module to obtain sentence-level embedding results and document-level embedding results for the event. The multi-granularity context-aware encoding module employs a two-level encoding method of intra-sentence self-attention aggregation and document-level cross-sentence attention weighting to capture the local sentence context and global document context of the event, respectively. An initial event causal graph structure is constructed. The initial event causal graph structure is encoded through the event causal graph encoding module to obtain the graph embedding result. The sentence-level embedding result, the document-level embedding result, and the graph embedding result are processed by the iterative graph optimization module to obtain the event causal relationship recognition result.

9. An event causal recognition device based on an iterative graph neural network, characterized in that, The device includes: a memory, a processor, and an event causality identification program based on an iterative graph neural network stored in the memory and running on the processor, the event causality identification program based on the iterative graph neural network configured to implement the steps of the event causality identification method based on an iterative graph neural network as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores an event causality recognition program based on an iterative graph neural network. When the event causality recognition program based on the iterative graph neural network is executed by a processor, it implements the steps of the event causality recognition method based on an iterative graph neural network as described in any one of claims 1-7.

Images