Sentiment Composition Using a Parabolic Model

Abstract

In this paper, we propose a computational model that accounts for the effects of negation and modality on opinion expressions. Based on linguistic experiments informed by native speakers, we distil these effects according to the type of modality and negation. The model relies on a parabolic representation where an opinion expression is represented as a point on a parabola. Negation is modelled as functions over this parabola whereas modality through a family of parabolas of different slopes; each slope corresponds to a different certainty degree. The model is evaluated using two experiments, one involving direct strength judgements on a 7-point scale and the other relying on a sentiment annotated corpus. The empirical evaluation of our model shows that it matches the way humans handle negation and modality in opinionated sentences