Description And Recognition Of Mongolian Object-predicate Relation Based On Dependency Grammar

The research of parsing is one of the key technology in Mongolian information processing.With the in-depth study of information processing,such as research and development in text proofreading and machine translation,the more accurate result of parsing is demanded in recent years.Mongolian grammatical study as theoretical analysis foundation,Mongolian lexical analysis and Mongolian dependency parsing as previous research achievements,this paper describes dynamic characteristic of contemporary Mongolian object-predicate relation and achieves automatic identification of it,using statistics and metrology.Object-predicate relation is a complex dependence relationship and it accounts for a large proportion in Mongolian sentence.Complex morphological change of Mongolian results in difficulty while accurately recognizing object-predicate relation.Accurately recognizing Mongolian object-predicate relation is significance in Mongolian parsing.The significance of accurately recognizing Mongolian object-predicate relation is as follows:for one thing,try to adopt a statistical means to traditional linguistics which introduces new experience in grammar research;for information processing part,the processed corpus has been expanded,at the same time,innovation model has been introduced to refinement of Mongolian parsing.This paper describes dynamic characteristics and study automatic identification of Mongolian object-predicate relations from following steps:First,expand,proofread and consummate contemporary Mongolian tree bank.A tree bank of 189048 words level,13154 sentences has been newly proofreaded.Second,detailed statistics for lexical feature,collocation feature and dependency syntax feature of Mongolian object-predicate relation provides theoretical basis for artificially programming rules and developing machine learning feature template.Third,four groups of experiments are designed to recognize Mongolian object-predicate relation:first,CRF statistical model based recognition experiment;then,CRF statistical model based recognition experiment after added artificially programmed rules;third,CRF statistical model based recognition experiment after added restricted rules;last,CRF statistical model based recognition experiment after added revised rules.The accuracy of four experiments is89.81%、89.80%、89.80% and 89.73% respectively.