Error Analysis for the Improvement of Subject Ellipsis Detection

Luz Rello , Gabriela Ferraro , Alicia Burga


This paper presents an analysis of the errors of a machine learning method that allow us to propose changes to improve it in future developments. The evaluated system detects Spanish subject ellipsis and yields an accuracy of 85.3%. We extract the wrongly classified instances of our training data (1,001) and classify the errors. We perform an analysis of these instances taking into account the features and the linguistic patterns involved, which motivate the inclusion of new features and rules in the system.

Texto completo: