A hierarchical perception based event extraction method and system

Abstract

The application relates to a kind of event extraction method and system based on hierarchical perception.The method comprises: inputting training data into event extraction module for event extraction, obtaining corresponding event type;Event type obtained by event extraction module is input into graph embedding module, graph embedding module learns event embedding in polar coordinates using causal relationship and sub-event relationship, and obtains event graph with hierarchical structure in rotation and modulus two dimensions;Event extraction module and graph embedding module are trained together, event embedding representation is enhanced by graph embedding module, and the performance of event extraction module is improved;Text to be input into the event extraction module of training completion is input into the event extraction module, and event extraction result is obtained.The application proposes HAEE model to construct event graph embedding, and the event extraction task under low resource condition can have higher performance, and embedding analysis in different dimensions proves that event graph embedding can effectively model event semantic hierarchical structure.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a hierarchical awareness-based event extraction method and system. Background Technology

[0002] Event extraction is an important task in natural language processing, aiming to extract structured event information from unstructured text to support downstream tasks such as question answering systems, information retrieval, and knowledge graph construction. For example, in the sentence "Stewart's first marriage to Alana Hamilton lasted about five years," the event extraction task needs to identify the event type "marriage," where the word "marriage" triggers the event.

[0003] In practical applications of event extraction, two main challenges exist. One is the problem of data scarcity. Event extraction requires a large amount of labeled data to train models. However, in real-world scenarios, the number of samples for certain event types is limited, which may hinder the model from effectively extracting event information. Therefore, it is necessary to explore techniques that can learn knowledge from limited or even missing data. On the other hand, due to the development of large-scale language models, learning existing knowledge is no longer a difficult problem. Instead, constructing knowledge with semantic features is a more challenging problem. Training models to capture deep semantic associations and extract latent information can enhance the performance of event extraction models. Therefore, it is necessary to study hierarchical awareness methods for event extraction tasks.

[0004] In low-resource event extraction scenarios, using more abstract event relationships as additional knowledge is a promising direction. By learning event-event relationships, a higher level of event understanding can be provided, which can be used to build better event representations and guide event extraction models. OntoED (Deng,S.,Zhang,N.,Li,L.,Hui,C.,Huaixiao,T.,Chen,M.,Huang,F.,Chen,H.:OntoED:Low-resource event detection with ontologyembedding.In:Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1:Long Papers).pp.2828–2839.AssociationforComputationalLinguistics,Online(2021).) is a classic and important model that formalizes event extraction as an event ontology task. By introducing three types of relationships (time, cause, and sub-event), it establishes connections between events, directing information towards low-resource events. To some extent, it addresses the data scarcity problem. However, OntoED still has some limitations. Event relationships can coexist and interact simultaneously in time and space; however, OntoED only considers and computes each relationship individually, without discussing whether interactions exist between them. Furthermore, OntoED only deals with first-level sub-events, without considering whether it has multi-level features, which may lead to the loss of relational knowledge.

[0005] In low-resource scenarios, leveraging stronger models and data augmentation are two main directions for improving event extraction tasks. Leveraging stronger models involves designing models that enhance their learning capabilities, thereby making full use of limited samples and reducing their dependence on data. Data augmentation aims to improve sample quality and provide more semantic representations. Compared to traditional machine learning methods that struggle to handle complex event structures and semantic relationships, most recent event extraction work is based on neural networks that can automatically learn useful features from data. Existing RNN and CNN-based models only consider the semantic information of individual samples without considering their relationships with other samples. Pre-trained models have attracted widespread attention and achieved rapid development due to their powerful text learning capabilities. In event extraction tasks, the focus has shifted from building lexical representations for trigger identification to representing samples as a whole. However, these techniques still face challenges, such as the need for large amounts of labeled data and the inability to explore more suitable strategies. Another research direction for improving model performance is through data augmentation. However, existing techniques mainly focus on augmenting samples at a local level without improving sample quality from a higher-level perspective. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a hierarchical awareness-based event extraction method and system.

[0007] The technical solution adopted in this invention is as follows:

[0008] A hierarchical-aware event extraction method includes the following steps:

[0009] The training data is input into the event extraction module to extract events and obtain the corresponding event types;

[0010] The event types obtained from the event extraction module are input into the graph embedding module. The graph embedding module learns the event embedding in polar coordinates using causal relationships and sub-event relationships, resulting in an event graph with a hierarchical structure in both rotation and modulus dimensions.

[0011] The event extraction module and the graph embedding module are trained together. The graph embedding module enhances the event embedding representation and improves the performance of the event extraction module.

[0012] The text to be extracted is input into the trained event extraction module to obtain the event extraction results.

[0013] Furthermore, the graph embedding module constructs event pairs with sub-event relationships at different layers using modulus, and constructs event pairs with causal relationships at the same layer using rotation. In this way, it makes full use of the interaction between different relationships, thereby enhancing the event embedding representation.

[0014] Furthermore, the graph embedding module divides the embedding v of event type e into rotational embedding v of equal size. r Modular embedding v m and obfuscated embedding v mix By integrating features from both rotation and modulus dimensions through the obfuscated embedding, the generalization ability of event representation is enhanced; [v r ,v mix Mapping [v] to rotation as a "causal" representation m ,v mix Mapping to a modulus as a "sub-event" representation enhances the interaction between the "causal" and "sub-event" relationships, helping to discover potential relationships between events and ultimately enhancing the model's semantic understanding capabilities.

[0015] Furthermore, during the training of the event extraction module and the graph embedding module together, convolution operations are performed on the event embeddings to enhance the representation of events.

[0016] Furthermore, the joint training of the event extraction module and the graph embedding module is achieved by balancing three loss functions, including the event extraction loss function. Rotation loss function Modular loss function The final loss function is expressed as:

[0017]

[0018] Where σ1, σ2, and σ3 are learnable parameters.

[0019] A hierarchical event extraction system, comprising:

[0020] The event extraction module is used to extract events from the input training data to obtain the corresponding event types; and after training is completed, to extract events from the input text to be extracted to obtain the event extraction results.

[0021] The graph embedding module is used to learn event embedding in polar coordinates based on the event types obtained by the event extraction module, using causal relationships and sub-event relationships, to obtain an event graph with a hierarchical structure in both rotation and modulus dimensions.

[0022] The training module is used to train the event extraction module and the graph embedding module together. The graph embedding module enhances the event embedding representation and improves the performance of the event extraction module.

[0023] The beneficial effects of this invention are as follows:

[0024] Event extraction aims to extract structured event information from unstructured text. Most recent work relies heavily on large labeled training sets and lacks semantic construction of events, thus remaining dependent on data. This invention proposes the HAEE model to construct event graph embeddings. While learning the mapping from samples to events, it models events on two dimensions of polar coordinates through causal and sub-event relationships, thereby achieving the goal of distinguishing events and establishing event-to-event relationships in graph embedding. Specifically, event pairs with sub-event relationships are constructed at different layers using moduli, while event pairs with causal relationships are constructed at the same layer through rotation. This approach fully utilizes the interactions between different relationships, thereby enhancing the event embedding representation. Experiments show that HAEE achieves high performance in event extraction tasks under low-resource conditions, and embedding analysis across different dimensions demonstrates that event graph embeddings can effectively model the semantic hierarchy of events. Attached Figure Description

[0025] Figure 1 An event graph contains causal relationships and sub-event relationships. In an event graph, the underlying relationships between "causality" and "sub-events" may not be visible in the knowledge set, but they are real and can be uncovered. The graph illustrates events that are in the same time and space and are related through "sub-events".

[0026] Figure 2 : A schematic diagram of the composition of the HAEE model of the present invention. Detailed Implementation

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0028] To address the problems in existing technologies, this invention establishes a knowledge set encompassing the relationships between causes and sub-events, considering their correlation characteristics. This invention does not consider time, as it is relatively weak in terms of logical relevance. Figure 1 As shown, a causal relationship indicates that the head event leads to the tail event, while a sub-event relationship indicates that the sub-event is a component of the parent event. Note that due to the transitivity of sub-events, it can be a multi-level relationship. By modeling both causes and sub-events simultaneously, we can observe the coexistence and interaction between relationships: events occurring in the same time and space may influence events occurring in another time and space. Establishing cause-sub-event relationships not only makes the correlations closer but also helps to explore the potential connections between events.

[0029] This invention proposes a novel hierarchical awareness model called HAEE (Hierarchy-Aware with Event Graph Embedding), which combines cause-sub-event relationships and event graph embedding. HAEE aims to distinguish events and establish relationships between them at the same and different levels of the hierarchy through rotation and modular operations. For causes, this invention employs a rotation-based approach, where event pairs are positioned opposite each other on a circle. For sub-events, this invention uses a contrastive learning method, where the distance between sub-events is as close as possible to the distance from the sub-event to its parent node. Furthermore, convolutional operations are performed on the embeddings to enhance the representation of events. In this way, further and hierarchical relationship knowledge between events can be learned. This invention uses an uncertainty-weighted loss function to compute multiple loss functions to obtain the final loss, which helps to balance the importance of each loss in the final event extraction task.

[0030] By applying these ideas, HAEE can start with a small number of labeled samples and gradually discover more potential knowledge contained in event graphs (such as...). Figure 1 (As shown by the dashed line in the middle), thereby improving the performance of low-resource event extraction tasks. Experiments show that HAEE outperforms existing models in situations where data is scarce. To further analyze the effectiveness of the model of this invention, a modular analysis was conducted to examine each component of the model and its contribution to the overall performance. This invention found that the model of this invention has good modeling and semantic hierarchy in event embedding, which is a key factor in accurately extracting events from text.

[0031] In summary, the contributions of this invention are as follows:

[0032] 1) This invention proposes a novel hierarchical perception model called HAEE. By constructing a knowledge set containing "causes" and "sub-events", it can discover more potential information and allow knowledge to flow into low-resource events to solve the event extraction problem in low-resource scenarios.

[0033] 2) In order to make full use of the coexistence and interaction between time and space relationships, this invention utilizes the "cause-sub-event" relationship to learn event embedding in polar coordinates, which constitutes an event graph with a hierarchical structure in the rotation and modulus dimensions.

[0034] 3) Experimental results show that HAEE achieves better performance in low-resource event extraction tasks. This invention conducted a modular analysis, examining each component of the model and its contribution to overall performance, demonstrating that the model of this invention has good modeling and semantic hierarchy on the event graph.

[0035] The HAEE model proposed in this invention is described in detail below. This invention first introduces event extraction as the main task of the model. Next, it introduces graph embedding, including rotation and modulus components, for hierarchical awareness event graph modeling. Finally, it introduces embedding convolution and loss functions to complete the remaining parts of the model.

[0036] 1. Problem Modeling

[0037] Figure 2 This illustrates the detailed processing of the HAEE model, including event extraction, the rotation component in graph embedding, and the modulus component in graph embedding. The model comprises three steps: Step 1 (Graph Modeling) forms sample representations and models event embeddings as a graph. Step 2 (Score Mapping) calculates event probabilities, rotation distances, and modulus distances (or modular distances). Step 3 (Loss Calculation) forms three loss functions and combines them to obtain the final loss.

[0038] For an event extraction task, the given input includes a sample set T = {X} i |i=1,...,N} and a set of event types ε={e i |i=1,...,M}. Each sample X i In T, there is a sequence of tokens {x}. 1 ,...,x L}, where L is the maximum length of the sequence, and k is the position of the trigger in the sequence. The event type set ε contains different event types e, each X i Belongs to an event type tag e i The goal of the event extraction task is to predict the event type label for each sample.

[0039] Event relationships refer to the more abstract connections between events, where the set of relationships R includes the set of causal relationships R0. c The set of sub-event relations R s The set of causal relationships R c Including causing r cause and caused by r causedby The set of sub-event relationships R s Including sub-events r subevent And the father event r superevent Based on the event type set ε and the relation set R, a knowledge set K about events and relations is constructed, consisting of triples (e...). h ,r,e t )∈ε×R×ε constitutes.

[0040] 2. Model Overview

[0041] This invention proposes a general model called HAEE, which includes three modules: event extraction, graph embedding, and embedding convolution. Key information for each module is as follows: Figure 2 As shown, graph embedding consists of a rotation part and a modular part. Each part, including event extraction, can be divided into three steps: graph modeling, score mapping, and loss calculation.

[0042] For event extraction, this invention obtains the context representation of a sample from BERT and calculates the probability between the sample and its event to form the event extraction loss.

[0043] For graph embedding, this invention models event embedding in polar coordinates, finds events in the knowledge base that are "causally" or "sub-event" related to the query event, and calculates the corresponding distance scores to form a graph embedding loss function.

[0044] For embedded convolution, this invention learns more advanced and hierarchical relational knowledge by convolving distant events.

[0045] The overall process of the hierarchical awareness-based event extraction method of the present invention is as follows:

[0046] A) Training Phase:

[0047] The training data is input into the event extraction module to extract events and obtain the corresponding event types;

[0048] The event types obtained from the event extraction module are input into the graph embedding module. The graph embedding module learns the event embedding in polar coordinates using causal relationships and sub-event relationships, resulting in an event graph with a hierarchical structure in both rotation and modulus dimensions.

[0049] The event extraction module and the graph embedding module are trained together. The graph embedding module enhances the event embedding representation and improves the performance of the event extraction module.

[0050] B) Practical Application Stage:

[0051] The text to be extracted is input into the trained event extraction module to obtain the event extraction results.

[0052] 3. Event Extraction

[0053] To obtain each sample X i This invention utilizes a pre-trained BERT model. Specifically, it takes the average of the word embeddings from the first and last layers of the BERT transformer to form the sentence representation. This method preserves the semantic information of words and sentences in the original samples to the greatest extent possible, ultimately enhancing the event representation.

[0054] This invention uses the same formula as OntoED to calculate event probabilities. For a given token sequence X... i It belongs to event type e k The probability is:

[0055]

[0056] Where, P(y=e k ) indicates that the event type y of the current sample belongs to e. k The probability, v j v represents the event type traversed in each iteration when summing and traversing the event type set. k ε represents the type of event being computed, and ε represents the set of types mentioned in the problem modeling.

[0057] The final event extraction loss function is:

[0058]

[0059] Among them, e q Indicates the actual event type of the current sample.

[0060] 4. Image Embedding

[0061] Based on the characteristics of "causality" and "sub-events," this invention chooses two dimensions of polar coordinates to model them. Polar coordinates are a position system in two-dimensional space, consisting of rotation and modulus. By determining the rotation and modulus of a point, its position in two-dimensional space can be determined.

[0062] To map event types to polar coordinates, this invention divides the embedding v of event type e into equally sized rotational embeddings v. r Modular embedding v m and obfuscated embedding v mix The addition of obfuscated embeddings aims to synthesize features from both dimensions, enhancing the generalization ability of event representations. Specifically, [v r ,v mix Mapping [v] to rotation as a "causal" representation, [v] m ,v mix This is mapped to a modulus as a representation of "sub-events". This modeling approach enhances the interaction between two relations, helps to discover potential relationships between events, and ultimately enhances the model's semantic understanding capabilities.

[0063] 4.1) The rotational part of "causality" refers to the occurrence of the head event leading to the occurrence of the tail event, indicating a clear causal relationship between them. This is demonstrated by the causal set R. cThis invention models events as rotations on polar coordinates. Specifically, to represent the "causal" relationship between two events, this invention learns their relative positions on a circle.

[0064] Given a query event e and a "cause" r c Find all sets of knowledge that satisfy (e, r) in the knowledge set K. c ,e c ) event e c This forms a set of event types ε c For event types e and e c Embedded v and v c ,have:

[0065] (vv c (3) mod 2π = π

[0066] Where v, v c ∈[0,2π). The distance function is defined as:

[0067] D(e h e t )=||sin((v h -v t +π) / 2)||1. (4)

[0068] Among them, v h and v t These represent event type e respectively. h and e t Embedded.

[0069] Rotational loss is defined as:

[0070]

[0071] Where σ(·) represents the sigmoid function, and λ is a fixed threshold.

[0072] 4.2) Modular part

[0073] A "sub-event" refers to a situation where the head event contains a tail event in both time and space, where the tail event is a component of the head event. For the set R of "sub-events"... s Assuming that "sub-events" have multi-level characteristics, this means that events at the same level should have similar attributes, and there is a certain distance between an event and its parent. "Sub-events" are transitive, i.e., (a,r,b), (b,r,c), then (a,r,c). This invention models this as a representation of event levels on a polar coordinate modulus. To achieve this goal, contrastive learning is used to make events at the same level closer together and events at different levels further apart.

[0074] This invention first bases the query event e and the sub-event r on... s Find the set of knowledge K that satisfies (e, r) s ,e s All events e s And use them to form the event type set ε s Then, this invention defines two sets of event pairs: P d Contains pairs (e,e) at different levels i ), where e i ∈ε s P s Includes pairs at the same level (e) i ,e j ), where e i ,e j ∈ε s And i≠j.

[0075] The distance between events is defined as the Euclidean distance between their corresponding vectors:

[0076] D(e h e t )=||v h -v t ||2. (6)

[0077] Finally, this invention defines the modulus loss as:

[0078]

[0079] Where γ represents the interval threshold under the modulus.

[0080] 5. Embedded Convolution

[0081] For an event relation chain (a,r,b), (b,r,c)∈K, and Event c might play a role in learning the relation (a, r, b) of event a. To enhance the representation of events and gain more hierarchical knowledge for relation learning, given a query event e, relation r, and event relation knowledge (e, r, e...),... r In the case of ), the present invention finds all that satisfy (e) r ,r,e * Event e in knowledge set K * To form an event set ε * Then calculate e. r The embedding representation, i.e., embedding convolution, is:

[0082]

[0083] Where v *This is the embedding representation of e*, where α∈[0,1] are the weights of the remote events. This method yields richer and more accurate event representations. This embedding convolution is used in the graph embedding module.

[0084] It should be noted that the embedding convolution in this invention is a further improvement and optimization step, not a necessary step. That is, good results can be achieved without embedding convolution.

[0085] 6. Loss Function

[0086] Within each event where relation calculations are performed, the distribution of relations may be uneven, resulting in some events having many relations while others have few. To balance the three types of losses, this invention introduces an uncertainty-weighted loss multi-task optimization strategy. By incorporating weight coefficients into the loss and adjusting their weights according to the prediction uncertainties of different modules, this invention can express the final loss function as:

[0087]

[0088] Where σ1, σ2, and σ3 are learnable parameters.

[0089] 7. Experiment

[0090] In this section, the invention first introduces two datasets (OntoEvent and MAVEN-Few) and benchmark methods. Then, the invention reports experimental results for the event extraction task in different contexts, including a) overall evaluation and b) low-resource evaluation.

[0091] 7.1 Dataset

[0092] This model evaluation used two datasets: the OntoEvent dataset and a selected MAVEN-ERE dataset, referred to as MAVEN-Few. OntoEvent is an event dataset with event relationships, used to demonstrate that introducing ontology facilitates event extraction. MAVEN-ERE is a unified large-scale event relationship extraction dataset based on the original MAVEN dataset, containing a large number of event types and their relationships. For ease of comparison with OntoED, 71 event types present in OntoEvent and their corresponding samples were selected from MAVEN-ERE to construct the MAVEN-Few dataset for evaluation. To construct the knowledge set, all relevant event relationships were selected from MAVEN-ERE, including 277 "causes" and 83 "sub-events" for experiments. Compared to OntoEvent's 9 "causes" and lack of multi-level "sub-events," the relationships in MAVEN-Few are richer, potentially making it easier to uncover more potential associations between events. In this invention, these two datasets and the knowledge set will be used to evaluate the performance of HAEE.

[0093] Table 1 shows the statistics for the OntoEvent and MAVEN-Few datasets, where: Doc: document, Train: training set, Valid: validation set, Test: test set, Class: event type, Caus-Rel: causal relationship, Sub-Rel: multi-level sub-event relationship.

[0094] Table 1: Statistics of OntoEvent and MAVEN-Few datasets

[0095]

[0096] 7.2 Experimental Setup

[0097] This invention tested the validation set results across different dimensions {50, 100, 200, 500}, and ultimately selected 100 as the dimension for event embedding. The maximum token sequence length for the training samples was set to 128, and a dropout ratio of 0.2 was used to prevent overfitting. The learning rate was 1×10⁻⁶. 3The initial values ​​of the uncertainty values ​​σ1, σ2, and σ3 are set to -0.5. Regarding hyperparameter selection, this invention uses a grid search on the validation set. Specifically, it searches for optimal values ​​for convolution weights α∈[0.2, 0.8] with a stride of 0.2, rotation thresholds λ∈[0.06, 0.12] with a stride of 0.02, and modulus thresholds γ∈[0.06, 0.12] with a stride of 0.02 on the validation set. The final values ​​used for testing are set to 0.4, 0.08, and 0.08, respectively. This invention randomly divides the dataset into a training set (80%), a validation set (10%), and a test set (10%). Using the SGD optimizer, the batch size for each training iteration is 42 samples, and a total of 5000 iterations are performed to obtain the final result. This invention uses precision, recall, and F1 score to evaluate the model's performance in classifying samples to the correct event labels.

[0098] 7.3 Reference Model

[0099] For the overall evaluation of the event extraction task, this invention compares the proposed model with the following baseline methods:

[0100] AD-DMBERT is an adversarial training model that enhances remotely supervised event detection models and automatically builds more diverse and accurate training data for semi-supervised event detection models.

[0101] OneIE is a joint information extraction neural model. It explicitly models interdependencies across subtasks and across instances, and predicts outcomes as a unified graph rather than isolated knowledge elements.

[0102] PathLM is an autoregressive language model that aims to induce graph patterns by connecting two event types through multiple paths involving entities.

[0103] OntoED describes the event extraction task as an event ontology task and infers richer samples and event-event relationships through ontology learning.

[0104] 7.4 Experimental Results

[0105] a) Overall Assessment

[0106] Table 2 shows that on the OntoEvent dataset, HAEE outperforms other models (such as BERT-based AD-DMBERT) across all three metrics. This indicates that the BERT-based HAEE framework can better establish connections between events. It also outperforms graph-based OneIE and PathLM, which only convert sentences into instance graphs and ignore potential relationships between events. Ontology-based OntoED, however, only explains event relationships at the semantic level, without considering whether interactions exist between relationships, and its modeling of "cause" and "sub-event" is weak.

[0107] Table 2: Evaluation of event extraction from the entire OntoEvent dataset.

[0108] Model Precision Recall F1 Score AD-DMBERT 0.6735 0.7346 0.7189 OneIE 0.7194 0.6852 0.7177 PathLM 0.7351 0.6874 0.7283 OntoED 0.7756 0.7844 0.78 HAEE 0.8882 0.8868 0.8875

[0109] b) Low resource assessment

[0110] Table 3 shows that HAEE maintains better performance in low-resource scenarios. When the training samples in the OntoEvent dataset are reduced to 50%, OntoED's F1 score drops by 6.5%, while HAEE's drops by only 1.3%. When the training set is reduced to 25%, OntoED's F1 score drops by 16%, while HAEE's drops by only 2.4%. Even when the training set is reduced to 10%, HAEE only drops by 7.6%, with an F1 score of 0.831, significantly higher than OntoED's F1 score of 0.78 on the full training set.

[0111] Table 3: Event extraction F1 scores on OntoEvent and MAVEN-Few training data at different scales.

[0112]

[0113] On the low-resource MAVEN-Few dataset, HAEE still performs well. Its F1 score is higher than OntoED in every scenario. Regarding model stability, OntoED shows a significant drop when the training set is reduced to 50%, while for HAEE, the F1 score only drops by 2.5% and 5.6% when the training set is reduced to 50% and 25%, respectively. Only when it is reduced to 10% does the F1 score show a significant drop of 27%. However, compared to OntoED, HAEE still has a significant advantage at a data ratio of 25%, with an F1 score 4.4% higher than OntoED on the full dataset. It can be seen that HAEE can maintain good and stable performance in low-resource scenarios, while OntoED relies solely on semantic relationships. This means that explicit event relationships are insufficient to support event representation in low-resource scenarios because it does not consider implicit interactions between relationships. Since HAEE models events on a graph, it can fully utilize event relationships to enhance event representation in low-resource scenarios, thereby helping the model perform better in event extraction tasks and improving its robustness.

[0114] 8. Analysis

[0115] 8.1 Embedding Relevance

[0116] To analyze the event embedding as a whole, this invention selects several events from the event embeddings trained by the model to calculate the correlation between them. By embedding events into the graph, the model can learn the relationships between causality and the interaction of sub-events.

[0117] Similarly, this invention also allows us to observe that the model can learn the relevance of similar words, proving that the model can ultimately enable event embeddings to have a certain semantic capability.

[0118] 8.2 Ablation Experiment

[0119] To investigate the impact of each module in the model, this invention removed the modulus and rotation components on the OntoEvent dataset under different resource scenarios and tested only using the basic architecture of the model. As shown in Table 4, when there is sufficient data, the improvement in accuracy by each module is not significant. This may be because the sample set itself already contains enough information, making it difficult to demonstrate the effects of rotation and modulus.

[0120] However, under limited resources, the combination of rotation and modulus can effectively improve the F1 score. It can be observed that under extremely low resource conditions (10%), using rotation alone can improve the F1 score by 2.2% compared to not using graph embedding. However, if the present invention combines rotation and modulus, the F1 score can be improved by 7.3%. Furthermore, when there is sufficient data, modulus provides more benefit, while under low resource conditions, rotation provides more benefit. By combining rotation and modulus, HAEE can fully leverage the relationships between events to enhance representational capabilities.

[0121] Table 4: The effects of different HAEE modules under low resource conditions.

[0122] Model Full 50％ 25％ 10％ HAEE 0.8875 0.8747 0.8634 0.831 HAEE w / o mod 0.8696 0.8593 0.8477 0.7808 HAEE w / o rot 0.8801 0.8641 0.8495 0.7578 HAEE w / o mod&rot 0.8741 0.8637 0.8474 0.7579

[0123] In summary, this invention proposes a novel Hierarchical Awareness Model (HAEE), which establishes "sub-event"-"causal" relationships through rotation and modulus, and maps event embeddings to polar coordinates to enhance event representation. To learn further and hierarchical relationship knowledge, this invention convolves distant events. This invention combines three loss functions by balancing the importance of each function to achieve better results in the final EE task. Experiments demonstrate that HAEE has high performance under low resource conditions, and the modeling effect of each module is analyzed, proving that the model has good modeling and semantic effects on event graphs.

[0124] Another embodiment of the present invention provides a hierarchical awareness-based event extraction system, comprising:

[0125] The event extraction module is used to extract events from the input training data to obtain the corresponding event types; and after training is completed, to extract events from the input text to be extracted to obtain the event extraction results.

[0126] The graph embedding module is used to learn event embedding in polar coordinates based on the event types obtained by the event extraction module, using causal relationships and sub-event relationships, to obtain an event graph with a hierarchical structure in both rotation and modulus dimensions.

[0127] The training module is used to train the event extraction module and the graph embedding module together. The graph embedding module enhances the event embedding representation and improves the performance of the event extraction module.

[0128] For the specific implementation process of each module, please refer to the description of the method of the present invention above.

[0129] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the method of the present invention.

[0130] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) storing a computer program that, when executed by a computer, implements the various steps of the method of the present invention.

[0131] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A hierarchical-aware event extraction method, characterized in that, Includes the following steps: The training data is input into the event extraction module to extract events and obtain the corresponding event types; The event types obtained from the event extraction module are input into the graph embedding module. The graph embedding module learns the event embedding in polar coordinates using causal relationships and sub-event relationships, resulting in an event graph with a hierarchical structure in both rotation and modulus dimensions. The event extraction module and the graph embedding module are trained together. The graph embedding module enhances the event embedding representation and improves the performance of the event extraction module. Input the text to be extracted into the trained event extraction module to obtain the event extraction result; The graph embedding module constructs event pairs with sub-event relationships in different layers using modularity, and constructs event pairs with causal relationships in the same layer using rotation. In this way, it makes full use of the interaction between different relationships, thereby enhancing the event embedding representation. The graph embedding module will classify event types. e Embedded v Divide into equally sized rotational embeddings v r Modular embedding v m and obfuscated embedding v mix By integrating the features of both rotation and modulus dimensions through the obfuscated embedding, the generalization ability of event representation is enhanced; [ v r , v mix Mapping [to rotation as a "causal" representation] will [ v m , v mix Mapping to a modulus as a "sub-event" representation enhances the interaction between the "causal" and "sub-event" relationships, helping to discover potential relationships between events and ultimately enhancing the model's semantic understanding capabilities. The event extraction module and graph embedding module are trained together by balancing three loss functions, including the event extraction loss function. Rotation loss function Modulus loss function The final loss function is expressed as: in, σ 1 , σ 2 , σ 3 is a learnable parameter.

2. The method according to claim 1, characterized in that, During the training of the event extraction module and the graph embedding module together, convolution operations are performed on the event embeddings to enhance the representation of events.

3. The method according to claim 1, characterized in that, The rotation loss function is: Among them, given query event e and causality r c In the knowledge set K Find all that satisfy ( e,r c , e c ) event e c This forms a set of event types. ; σ Represents the sigmoid function; λ It is a fixed threshold; D This is the distance function.

4. The method according to claim 1, characterized in that, The modulus loss function is: First, based on the query event e and subevents r s Find the knowledge set K The middle satisfies ( e,r s , e s All events e s This forms a set of event types. Then define two event pairs for the set. P d and P s ,in P d Contains pairs at different levels ( e,e i ), P s Includes pairs at the same level ( e i , e j ),and i ≠ j ; D It is a distance function; γ This represents the interval threshold under the modulus.

5. An event extraction system based on hierarchical awareness, characterized in that, include: The event extraction module is used to extract events from the input training data to obtain the corresponding event types; And after training, the event extraction results are obtained by extracting events from the input text to be extracted. The graph embedding module is used to learn event embedding in polar coordinates based on the event types obtained by the event extraction module, using causal relationships and sub-event relationships, to obtain an event graph with a hierarchical structure in both rotation and modulus dimensions. The training module is used to train the event extraction module and the graph embedding module together. The graph embedding module enhances the event embedding representation and improves the performance of the event extraction module. The graph embedding module constructs event pairs with sub-event relationships in different layers using modularity, and constructs event pairs with causal relationships in the same layer using rotation. In this way, it makes full use of the interaction between different relationships, thereby enhancing the event embedding representation. The graph embedding module will classify event types. e Embedded v Divide into equally sized rotational embeddings v r Modular embedding v m and obfuscated embedding v mix By integrating the features of both rotation and modulus dimensions through the obfuscated embedding, the generalization ability of event representation is enhanced; [ v r , v mix Mapping [to rotation as a "causal" representation] will [ v m , v mix Mapping to a modulus as a "sub-event" representation enhances the interaction between the "causal" and "sub-event" relationships, helping to discover potential relationships between events and ultimately enhancing the model's semantic understanding capabilities. The event extraction module and graph embedding module are trained together by balancing three loss functions, including the event extraction loss function. Rotation loss function Modulus loss function The final loss function is expressed as: in, σ 1 , σ 2 , σ 3 is a learnable parameter.

6. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 4.

Images