Document Level Subjectivity Classification Experiments in DEFT'09 Challenge
Cigdem Toprak, Iryna Gurevych · TUbilio (Technical University of Darmstadt) · 2009
In this paper, we present our supervised document level subjectivity classification experiments for English and French at the DEFT’09 Text Mining Challenge. We experiment with the word, POS, and lexicon-based features using an SVM classifier. Our word feature experiments (i) investigate the utility of the context information, and (ii) compare the binary and tf*idf feature representations in this task. We show that different class distributions favor different feature representations. Furthermore, on the English collection, we compare three, two of which are well-known, opinon lexicons at this task: the subjectivity clues from (Wiebe and Riloff, 2005; Wilson et al., 2005), SentiWordNet (Esuli and Sebastiani, 2006), and a list of verbs compiled from (Santini, 2007; Biber et al., 1999)1 . We show that, despite its limited coverage, the verb lexicon, consisting of 156 verbs, establishes relatively good results in English.