Prompt Based Event Argument Extraction

Event extraction is an important research direction in the field of information ex-traction,which aims to extract structured information from unstructured text.It is widely used in intelligence collection,knowledge extraction,document summarization,knowl-edge question answering and other fields.Event extraction handles various types of text,such as(online)news messages,blogs,and manuscripts.Among them,argument ex-traction is a sub-task of event extraction,which extracts corresponding arguments from sentences under the premise of given an event category and a corresponding event trig-ger.The current mainstream event argument extraction model is based on sequence label-ing.However,these methods cannot perform well under low data conditions,and these methods cannot effectively model the interaction between different arguments,let alone document-level cross-sentence extraction.In recent years,a method called cue learning has emerged.It uses the huge amount of parameters of the language model,and uses a piece of natural text as a prompt word to directly generate the results required by the task.This kind of method can solve the problem of too few training samples to a certain ex-tent.However,there is no effective method for constructing prompt sentences,and little is known about how to apply prompt learning to the argument extraction task.Therefore,this paper proposes a cue tuning method based on curriculum learning,studies how cue learning achieves excellent results under the condition of small samples,and proposes a method based on backdoor tuning to improve the composition of cue sen-tences.The contributions of this paper are as follows:· Aiming at the long-distance dependency and argument relationship modeling prob-lem faced by document-level event argument extraction.This paper proposes a method based on curriculum learning,which effectively models the argument re-lationship with the help of AMR graph,and effectively models the dependency between argument-argument and argument-trigger word through the method of cur-riculum learning.relation.State-of-the-art results are achieved on RAMS and Wiki Events datasets.· In order to further improve the construction of prompt sentences,we compared two prompt construction methods,name-based and ontology-based.The experimental results on three datasets show that the prompt sentences based on ontology can be extracted by relying on syntactic dependencies.to achieve better results.· Based on the above research,we analyze the argument extraction method based on cue learning from the perspective of causal analysis,find the backdoor path in it,and adjust it.The experimental results show that the method after backdoor adjustment is more efficient Robust and outperforms experimental settings with few samples.