| The explosive growth of network data brings challenges for people to obtain personalized semantic information.There is an urgent need for an automated information extraction technology to help people understand semantics.As a kind of automated information extraction technology,event extraction is a sub-task of information extraction,which aims to extract the events that people are interested in and annotating their semantic participants from a large amount of unstructured texts.It can be used in query intent analysis and automatic knowledge graph construction and so on.This paper aims to solve the cascading error,multi-label and ambiguity problems in Chinese and English event extraction.The main contents of this paper are as follows:(1)We propose to use classification training strategy to solve the problem of multilabel for event elements.Currently,most Chinese event extraction systems use a continuous pipeline model.It is prone to cascading errors,and downstream tasks can not feed back information to upstream tasks,assisting in the identification of upstream tasks.In this paper,event extraction is regarded as a sequence labeling task,and a joint Chinese event extraction model based on conditional random field(CRF)is constructed.Aiming at solving the multi-label problem of event elements in the joint model(That is,when an event contains more than one event,the same entity often plays different roles in different events).(2)A multi-task learning method is proposed to enhance each other's mutual learning under the same event category.Considering the event elements in event subclasses under the same event category have certain interrelatedness.To this end,this paper proposes a multi-task learning method for mutual enhancement of each event sub-class joint learning,and thus to some extent alleviates the imbalance problem of corpus.(3)We construct a recurrent neural network event detection model that uses document level information.The existing English event detection models mainly have two problems: 1)Although feature-based models use cross-sentence information,the need to carefully design a large number of manual features and inference rules.2)Although representation-based models can extract features automatically,it is difficult to accurately identify events using only the local sentence-level information.Therefore,we propose a document-level recurrent neural network event detection model.In order to validate the our assumptions,we conduct the experiments on Chinese and English corpora of ACE(Automatic Content Extraction)2005 respectively.Finally,the effectiveness of our method is proved by comparing with the-state-of-art methods. |
