A Linear Baseline Classifier for Cross-Lingual Pronoun Prediction
Jörg Tiedemann · 2016
This paper presents baseline models using linear classifiers for the pronoun translation task at WMT 2016.We explore various local context features and include history features of potential antecedents extracted by means of a simple PoSmatching strategy.The results show the difficulties of the task in general but also represent valuable baselines to compare other more-informed systems with.Our experiments reveal that the predictions of English correspondences for given ambiguous pronouns in French and German is easier than the other way around.This seems to verify that predictions, which need to follow more complex agreement constraints, require more reliable information about the referential links of the tokens to be inserted.