Multitask Item Response Models for Response Bias Removal from Affective Ratings

Response style (RS) is a tendency to choose specific categories regardless of content, e.g, extreme or midpoint categories. It degrades the validity of the analysis of subjective ratings such as correlation and variance-based analyses. However, the computational removal of RS has received little attention from the affective computing community. RS removal techniques have been proposed in areas such as marketing research. However, most of these techniques do not exploit the content-independence of RS; i.e. it should be observed consistently in various tasks, such as affective judgment tasks and standard psychological questionnaires. Therefore, this paper proposes a multitask RS removal method. An individual's responses in multiple tasks are modeled using task-independent RS parameters, and task-dependent parameters, including the item and respondent's characteristic parameters based on item response models (IRM). Through Bayesian modeling, we observed that: i) the proposed model outperformed traditional IRMs in terms of predictive accuracy; ii) our multitask framework estimated RS with higher precision than previous single-task-based RS removal methods; iii) our model replicated Japanese midpoint RS, which has been demonstrated repeatedly in previous cross-cultural studies; and iv) RS-removed predictive ratings showed higher inter-rater agreement than those including RS in valence/arousal judgment tasks.