Pronoun Language Model and Grammatical Heuristics for Aiding Pronoun Prediction

Abstract

The cross-lingual pronoun prediction task at WMT 2016 requires to restore the missing target pronouns from source text and target lemmatized and POS-tagged translations. We study the benefits for this task of a specific Pronoun Language Model (PLM), which captures the likelihood of a pronoun given the gender and number of the nouns or pronouns preceding it, on the target-side only. Experimenting with the English-to-French subtask, we select the best candidate pronoun by applying the PLM and additional heuristics based on French grammar rules to the target-side texts provided in the subtask. Although the PLM helps to outperform a random baseline, it still scores far lower than system using both source and target texts.