Dissertaties - Rijksuniversiteit Groningen
 
vp.jpg
English | Nederlands

Parse selection with Support Vector Machines

(2010) Borges, Francisco Dellatorre

Dit document is (ook) beschikbaar voor ruilverkeer - alleen door bibliotheken -. [Bestelformulier]


The goal of our research was to apply SVMs (Support Vector Machines) to the problem of parse selection. More specifically, to the parse trees produced by Alpino, and to compare its performance with the current Alpino disambiguation component, which is based on Maximum Entropy.
There are two basic problems to be dealt with when applying a machine learning technique to parse selection. The first is how to compare different parse trees to each other. We addressed this problem in the same way that it had been already addressed by Alpino, which allowed us to turn to (structured) trees into vectors in Nn.
The second issue is whether to consider the problem as a classification or as a regression problem. Many view parse selection as a classification problem, in a one-against-all manner, but it is actually a skewed regression problem. While the data modeled takes values in the interval [0,1], the evaluation of success is effectively measured by how well the upper end of the interval (the higher scores) were modeled. That holds as long as the lower scores are actually kept low somehow. Despite the skewness of the evaluation, we believe that parse selection is inherently a regression problem, and have chosen to use SV regression to perform parse selection, instead of the more popular SV classification.





Gebruik a.u.b. deze link om te verwijzen naar dit document:
http://irs.ub.rug.nl/ppn/324099886

Meer informatie in de catalogus
Meer informatie in Picarta



 
To top