EmotiWord: Affective Lexicon Creation with Application to Interaction and Multimedia Data

We present a fully automated algorithm for expanding an affective lexicon with new entries. Continuous valence ratings are estimated for unseen words using the underlying assumption that semantic similarity implies affective similarity. Starting from a set of manually annotated words, a linear affective model is trained using the least mean squares algorithm followed by feature selection. The proposed algorithm performs very well on reproducing the valence ratings of the Affective Norms for English Words (ANEW) and General Inquirer datasets. We then propose three simple linear and non-linear fusion schemes for investigating how lexical valence scores can be combined to produce sentence-level scores. These methods are tested on a sentence rating task of the SemEval 2007 corpus, on the ChIMP politeness and frustration detection dialogue task and on a movie subtitle polarity detection task.