Research On Target-dependent Event Extraction

Mining and screening interested entities and events from massive information texts on the Internet is not only an important research direction in the field of natural language processing,but also an important step in the construction of knowledge graph in the industry.Generally,event extraction focuses on events and explores the identification of event types and event elements.Most event-centric research methods detect events first and then extract event elements based on event patterns,which are limited by the performance of event detection.Based on the demand of extracting target entities and their corresponding event information in real industrial scenes,this paper proposes target-dependent event extraction from the perspective of entity,which includes target-dependent event detection and extends to event extraction,aiming at extracting entities and their corresponding event types or event argument roles in text.The first work of this paper focuses on target-dependent event detection,extracting entities from text and identifying the fine-grained event types corresponding to the target entities.In order to alleviate the problem of mismatching between entities and events,a target-dependent semantic and syntactic model is proposed in this paper.On the basis of entity extraction,we incorporate event trigger features and target-specific syntactic dependency distance into our model in the entity-level event detection stage.The experimental results on Fin News CN-8K show that the model can capture relative contextual words for a target entity and extract accurate and complete entity-event pairs even if in a multi-target or multi-event scenario.The corresponding research has been published in IEEE Big Data conference in 2021.Text semantic and syntactic structure not only shorten the distance between words,but also contain semantic and structural knowledge,which are helpful to capture the event information related to the target entity.Therefore,the second work of this paper proposes a semantic and structural graph-enhanced model for target-dependent event detection.Relying on sequence,semantic,syntactic dependency features and edgeenhanced graph convolution network,the model can further improve the performance of the sequence-based encoding method in event detection.The experimental result of integrating multiple graphs achieves the best performance,which verifies that multiple graph structures can complement each other to assist the target entity-level event detection.Relevant patent has been applied for this part of the work.Entities usually play a certain role in events.In order to further enrich event information and build a complete knowledge graph,the third work of this paper extends target-dependent event detection to target-dependent event extraction.To explore entity extraction and role recognition in corresponding events from the perspective of entity,we propose a semantic and structural graph-enhanced event extraction model which extracts entities and event trigger words first and then predicts the relationship between entities and event trigger words as the entity argument role.The experimental results on ACE 2005 show that our model outperforms previous methods in entity extraction and event detection.Besides,without relying on additional global features,argument role extraction integrating multiple graphs also achieves the best performance,which verifies the effectiveness of the graph-based encoding method in event extraction.This paper focuses on target-dependent event extraction.And we conduct experiments on Fin News CN-8K and ACE 2005 dataset by sequence-based and graphbased encoding methods respectively.The results show that our models can mine entity and event information more accurately,and provide entity and event elements for the construction of downstream knowledge graph.