| Event Argument Extraction(EAE for short)is subordinate to the research on event extraction,which aims to assign appropriate roles to candidate arguments in events.This research usually relies on annotated corpora,lacking in most languages because of the complicate annotating rules as well as time-consuming and laborious annotating work.Cross-lingual EAE can utilize relatively abundant corpora in the source language to build models,which can be applied directly to the target language side with scarce annotated data to extract corresponding information.Currently,this research is still in the exploratory stage and poses significant challenges.In this thesis,we investigate the mainstream methods for cross-lingual EAE,and further study the transferability features both in structure and semantics,and finally improve the performance of cross-lingual EAE by enhancing the representations of dependency tree as well as imposing semantic constraints between bilingual sentences.This thesis first provides a new universal EAE framework which can be compatible with multi-type encoders.In advance,there are another two main contributions as follows:(1)A method that enhances the representations of the consistency of dependency structures among different languages is proposed for cross-lingual EAE.It is different from the method of fusing the nodes of the entire dependency tree into the model to help transfer learning through distance filtering strategies,we find that in different languages,the nodes of the dependency path from trigger words to candidate arguments show high correlation in structure and degree of participation in events.Therefore,we introduce an additional BiGRU network to effectively model this crucial dependency path information and seamlessly integrate it into the main model..Experimental results show that the method can be flexibly integrated with mainstream cross-lingual EAE models,and effectively improves their performance.(2)A method of fusing similarity constraints between bilingual sentences is proposed for cross-lingual EAE.The incorporation of translation techniques in cross-lingual tasks is a common approach.However,in cross-lingual EAE tasks,the simple fusion of translated text into the model fails to achieve the expected results.In this paper,we delve into a novel approach by concatenating both the source text and its translated counterpart as model input.Furthermore,we introduce contrastive learning to enforce the similarity constraint of the sentence-level semantics between the two languages.This constraint aims to minimize the distance between positive example sentence pairs while maximizing the distance between negative example sentence pairs.Experimental results demonstrate the effectiveness of this method in significantly improving the performance of cross-lingual EAE.Remarkably,the performance of this proposed approach surpasses that of the method described in(1).Additionally,we combine the methods presented in(1)and(2),resulting in an optimal performance of the joint model,as evidenced by the experimental results. |
