Task-based Evaluation Report: Building a Dutch Subjectivity Lexicon

We describe a method for creating a Dutch subjectivity lexicon based on an English subjectivity lexicon, an online translation service and a Dutch general purpose thesaurus: Wordnet. We use a PageRank-like algorithm to bootstrap from the Dutch translation of the English lexicon and rank the words in the Dutch thesaurus by polarity. Two versions of the Dutch Wordnet are used in the experiments: the 2001 version and the 2008 version developed within the Cornetto project. We present the evaluation results based on human assessment of the top 2000 negative words and the top 1500 positive words in the resulting lexicons. We find that using Cornetto results in a 7% improvement in accuracy. Between 70% to 86% of this improvement can be attributed to the larger size of Cornetto, the remaining improvement is attributed to the larger set of relations between words.