Directed Graph Model for Mongolian Lexical Analysis

We propo se a gener ativ e stat istical model for Mongo lian lexical ana lysis. This model describes the lex ical analysis r esult as a dir ected g raph, wher e t he nodes r epr esent the stems, affixes and their tag s, while the edges represent the t ransitio n o r generation relationships between nodes. Especially in this wo rk, we adopt three kinds o f transitio n o r g eneration probabilities: a) pr obabilit ies of stem-stem tr ansition, affixaffix transitio n and stem-affix generation; b) t he transition o r generation probabilities between the co rr esponding tags; and c) the gener ation probabilities between stems o r affixes and their tags. U sing the 3rdlevel annotated corpus wit h about 200 000 w ords as the t raining data, this model achieves a wo rdlevel segmentation accuracy of 95. 1% , and a wo rdlevel jo int segmentat ion and tagging accuracy of 93% .