Abstract

This paper describes a model for the perceived emotion of Turkish sentences based on the emotions associated with the constituent words. In our model, each emotion is mapped to a point in the continuous space defined by three emotional attributes: valence, activation, and dominance. We collected a large data set through two independent surveys: a word-level survey that prompted users with emotional words and asked them to assign each word a continuous emotional interval, and a sentence-level survey that prompted users with emotional sentences collected from 31 children’s books and asked them to rate each sentence on a discrete emotional scale. The word-level survey was aimed at creating a core affective lexicon for Turkish. It is difficult to build a comprehensive affective lexicon for Turkish due to its very productive morphology that generates a very large vocabulary. We deal with the sparsity issues caused by the large word vocabulary by analyzing the emotional content of word roots. Our experimental results indicate that there is a strong correlation between the emotions attributed to Turkish word roots and the Turkish sentences.

Date

2012

Authors

Ozan Cakmak, Abe Kazemzadeh, Dogan Can, Serdar Yildirim, Shrikanth Narayanan

Journal

Corpora for research on emotion sentiment & social signals