Annotation and detection of conflict escalation in Political debates

Conflict escalation in multi-party conversations refers to an increase in the intensity of conflict during conversations. Here we study annotation and detection of conflict escalation in broadcast political debates towards a machine-mediated conflict management system. In this regard, we label conflict escalation using crowd-sourced annotations and predict it with automatically extracted conversational and prosodic features. In particular, to annotate the conflict escalation we deploy two different strategies, i.e., indirect inference and direct assessment; the direct assessment method refers to a way that annotators watch and compare two consecutive clips during the annotation process, while the indirect inference method indicates that each clip is independently annotated with respect to the level of conflict then the level conflict escalation is inferred by comparing annotations of two consecutive clips. Empirical results with 792 pairs of consecutive clips in classifying three types of conflict escalation, i.e., escalation, de-escalation, and constant, show that labels from direct assessment yield higher classification performance (45.3% unweighted accuracy (UA)) than the one from indirect inference (39.7% UA), although the annotations from both methods are highly correlated (ρ = 0.74 in continuous values and 63% agreement in ternary classes). Index Terms — Spoken Language Understanding, Conflicts, Paralinguistic, Spontaneous Conversation, Prosodic features, Turntaking features