| Event Extraction is a challenging task in Natural Language Understanding.It aims to extract specific event information from unstructured text,and the result is widely used in downstream fields such as Question Answering and Information Retrieval.The target of event extraction can further subdivided into three subtasks:Entity Mention Detection,Event Detection,and Argument Role Prediction.Previous work for Event Extraction is usually done in the form of pipelined approach.Recently,end-to-end models for Event Extraction are getting more attention,but two technical problems still exist.First,different subtasks focus on different granularities of context information,but in the existing methods,all three subtasks shared the representations from the same encoder,thus the information interested by a specific task might be lost.Second,triggers of different events from the same context has semantic relationship to each other,so do the argument roles of the same event.But in the existing methods,the interrelationships are limited by issues like time-dependency.Such limitations lead to insufficient use of the correlation between event elements in the sentence.Aiming at handling the above two problems,this thesis proposes an End-to-End Event Extraction framework Based on Multi-Grained Information Fusion.The framework consists of two main method,one is the Multi-Grained Dynamic Span Generator,and another one is the Span Ranking Network with Elements-Aware Attention for Event Extraction.The former method focuses on the first problem and assigns a task-specific span generator to fuse context features for a task from three different granularities:sentence,neighbor,and content.Capturing the features of interest for each subtask and meet their requirements for different granularities.While the latter model pays more attention to exploring the association of triggers-triggers and triggers-entities,and effectively models these relationships through a collaborative attention mechanism.The experiments on ACE-2005 prove that the proposed methods outperform the state-of-the-art baseline in Entity Mention Detection and Event Detection,while reaching a comparable result in Argument Role Prediction.The experiments verified the practicality and effectiveness of the proposed end-to-end Event Extraction methods from an application perspective. |
