Research On Event Detection Based On Entity Clues Mining

The task of event detection aims to automatically identify and classify event trigger words from unstructured text.Mining entity clues from raw language corpus can facilitate event detection because entity properties may imply the event type they belong to.Existing studies have introduced entity information by concatenating embedding representations,which is often based on local context modeling and cannot capture the rich semantics of entities.In addition,the correlation between entities and events has not been fully explored.To address these issues,this thesis proposes an entity clue mining-based event detection method,which explores entity clues from three perspectives:enhancing entity profiles,mining shared knowledge between entities and events,and constructing a global entity network.The pre-training process of language model representation only incorporates sentencelevel contextual information,resulting in a low coverage of entity profiles.Moreover,the thesis incorporates the interaction information of locally relevant entities.It also proposes a co-occurrence graph-based attention masking model to alleviate the interference of noise information during entity representation.The method is validated on the ACE-2005 corpus,demonstrating a 2.2%improvement in the F1 score compared to the baseline model.There is a strong correlation between events and entities,and shallow learning cannot efficiently capture the interaction information between them.This thesis proposes an approach that utilizes multi-task learning and a mixture of expert networks to automatically model the shared knowledge between entities and events.Firstly,It employs joint learning of event detection and entity recognition to automatically identify the relevance between entities and events.Secondly,It decouples and reconstructs the multiple features across tasks using independent "experts".The method achieves a 2.3%improvement in F1 score over the baseline model on the ACE-2005 dataset,and it outperforms state-of-the-art models to achieve the highest F1 score on the KBP-2017 dataset.Entities involved in events often span across multiple sentences,and the approach of mining entity clues using local context overlooks the global document information.To explore the correlation between docment-level entity mentions and events,this thesis constructs a global entity association network.It constructs entity candidate sets on a document level and employs a graph attention network to model the relationships between different candidate words in the target sentence and the entity candidate set.This method provides differentiated document-level entity information for different target words.It achieves a 3.3%improvement in F1 score over the baseline model on the ACE-2005 dataset.In conclusion,the proposed approaches in this thesis effectively exploit entity clues to improve the performance of event detection.