Research On Chinese Event Argument Extraction

Nowadays, most of the event extraction researches are based on English corpus whileChinese event extraction is still at an elementary stage. As a topic-based language, theexpression of Chinese sentences is very flexible, which leads to the loose connectionbetween the arguments and the trigger in an event. Due to the syntactic features based onshallow semantic are widely used in the previous studies, it leads to the low performanceof Chinese argument extraction.To address this issue, we propose a new method and more effective features forChinese argument extraction. The study can be concluded as follow.Firstly, the previous studies mainly use syntactic features which centered on thepredicate of the sentence to extract the arguments of the event. Unlike verbal trigger,nominal trigger does not act as the predicate in the sentence, which leads to the lowperformance of the argument extraction of the sentences which have a nominal trigger. Toaddress this issue, we propose a novel argument extraction approach, which base on triggermapping. It maps the nominal anchor to its predicate and uses it as a key to extractsyntactic features for classification. Experimental results on ACE2005Chinese Corpusshow that, in terms of F1-measure in argument detection and identification, our approachcan be improved by2.5%and2.8%respectively, compared with the baseline.Secondly, Chinese does not have a strict syntactic structure and ellipsis is a commonphenomenon, which leads to weaken the function of syntactic structure in argumentextraction. To address this issue, we make use of the semantic of roles, entities and triggerto be a supplement of syntactic-based approaches. Compared with the state-of-the-artsystem, the F1-measure of our approach can be improved by3.7%and4.2%respectively.Finally, the previous studies mainly use the sentence-level features to extract thearguments, but it is not effective to extract the arguments which have a long distance with the trigger. To address this issue, we propose a novel argument extraction approach whichbase on inference. This approach makes use of the consistency of cross-event argumentrole and gets through the known event mentions to extract the event mentions which lackof effective information. Experimental results on ACE2005Chinese Corpus show that, forAttack, Die and Injure events, in terms of F1-measure in argument detection andidentification, our approach can be improved by3.4%and3.7%respectively, comparedwith the baseline.