Multidimensional Transformation-Based Learning

Abstract

This paper presents a novel method that al-lows a machine learning algorithm following the transformation-based learning paradigm (Brill 1995) to be applied to multiple classification tasks by training jointly and simultaneously on all fields The motivation for constructing such a system stems from the observation that many tasks in natural language processing are naturally composed of multiple subtasks which need to be resolved simultaneously; also tasks usually learned in isolation can possibly benefit from being learned in a joint framework, as the signals for the extra tasks usually constitute inductive bias. The proposed algorithm is evaluated in two exper iments: in one, the system is used to jointly predict the part-of-speech and text chunks/baseNP chunks of an English corpus; and in the second it is used to learn the joint prediction of word segment boundaries and part-of-speech tagging for Chinese. The results show that the simultaneous learning of mul tiple tasks does achieve an improvement in each task upon training the same tasks sequentially. The part of-speech tagging result of 96.63% is state-of-the-art for individual systems on the particular train/test split.