Machine Learning of Language: A Model and a Problem

I will show how the memory-based approach, developed at the University of Tilburg and the University of Antwerp, provides an elegant and efficient model that fits natural language processing (NLP) tasks ranging from low-level problems like part-of-speech tagging to high-level applications like information extraction from text. Applications of this model lead to competitive accuracies compared to alternative approaches, suggesting that the proposed learning method has "the right bias" for NLP tasks. However, in the second part of the talk, I will show that the superiority of one learning approach over another for NLP tasks cannot be shown in a reliable way using current methodology for comparative machine learning experiments. I will go into the consequences of and possible patches for this methodological problem which affects a large part of current research in machine learning of language.
In case of problems or for comments, please contact: kuebler@sfs.uni-tuebingen.de ( Last modified: Fri Jun 14 18:26:50 MET DST 2002 )