Research On Sentence-Level Event Extraction Based On Information Fusion And Adversarial Prompt

Event extraction is one of the important research topics in the field of information extraction,which goal is to automatically extract structured event semantic information from unstructured textual data.It helps to enhance event-centric natural language understanding and generation.Due to the continuous growth of unstructured textual data on the Internet,there is an urgent need for more effective ways to automatically discover and extract specific types of event knowledge from text.As a result,event extraction has received significant attention from academia and industry in recent years.However,due to the complexity of event structures,event extraction techniques still face many challenges.One of the crucial challenges is obtaining important event-related information from within the text or from large-scale pre-trained models.This study focuses on the two core tasks of sentence-level event extraction,namely event detection and event argument extraction.It systematically explores event extraction techniques based on information fusion and adversarial prompts.The main research work is as follows:(1)For the event detection task,existing approaches leverage syntactic relations to enhance event detection.However,they may confuse different syntactic relations and introduce redundancy or noise information,which may lead to performance degradation.To this end,this thesis proposes a new event detection model based on information fusion named Dual GAT,which comprehensively exploits the complementary nature of syntactic and semantic relations to alleviate the above problems.Specifically,the Dual GAT model first constructs a dual relational graph that both aggregates syntactic and semantic relations to the key nodes in the graph,so that event-relevant information can be comprehensively captured from multiple perspectives.Then,it adopts augmented relational graph attention networks to encode the graph and optimize its attention weights by introducing contextual information,which further improves the performance of event detection.Extensive experiments conducted on the standard ACE2005 benchmark dataset indicate that Dual GAT significantly outperforms the baseline models involved in the comparison and verifies the superiority of Dual GAT over existing syntactic-based methods.(2)For the event argument extraction task,existing approaches leverage prompt templates and pre-trained language models to jointly enhance argument extraction capability.However,manually designed prompt templates usually have certain subjective limitations,and recent soft prompt-based methods are sensitive to model initialization and training methods,which may easily introduce external noise or bias information.To this end,this thesis proposes a new event argument extraction model based on adversarial prompt named ADST,which utilizes shared knowledge across different event types to alleviate the above problems.Specifically,ADST constructs a joint prompt by adding typespecific discrete text to the soft prompt.This reduces the difficulty of initializing the soft prompt and enables the prompt templates to have type-specific knowledge.Then,it utilizes task-aware and adversarial argument extractors to jointly train the constructed prompt templates.The ADST model can improve the quality of prompt templates through shared and type-specific argument knowledge,which effectively improves the performance of event argument extraction.Extensive experiments conducted on the standard ACE2005 benchmark dataset indicate that ADST significantly outperforms the baseline models involved in the comparison and verifies the effectiveness of ADST in low-resource scenarios.