Simultaneous feature selection and parameter optimization for memory-based natural language processing
Anne Kool, Jakub Zavrel, Walter M. P. Daelemans · 2000
We investigate the eects of (i) feature subset selection, (ii) parameter optimization, and (iii) simultaneous feature selection and parameter optimization in Memory-based Natural Language Processing (MBLP). We use a simple genetic algorithm for this problem and compare it to two iterative search methods on some typical tasks in natural language processing: partof -speech tagging of known and unknown words, and grapheme to phoneme conversion with stress assignment. We nd that (i) feature selection always outperforms the MBLP variant without selection, (ii) optimization of parameters for each specic task is benecial, and (iii) the combination of parameter optimization and feature selection performed simultaneously can lead to more accurate classication results. However, we have found no indications on our data that gas reach a signicantly better accuracy than the iterative methods, and, in general, the approach promises larger gains for more eective search methods.