홈 › Papers

Emotion Flip Reasoning in Multiparty Conversations

Emotion plays a major role in conversation analysis. To understand the emotional dynamics of the speakers in a conversation, it is imperative to reason out any change/flip of the emotion of a particular speaker. In this paper, we introduce a novel task -- Emotion Flip Reasoning (aka EFR) that aims to find a rationale or instigator behind the emotion flip of a speaker within a conversational dialog, e.g., joy --> anger. To this end, we introduce MELD-I, a new dataset with ground-truth EFR instigator labels. Following emotion psychology, we define 27 instigator labels in a hierarchical fashion - the coarse-grained set contains 14 instigators, while the fine-grained set comprises all 27 instigators. To benchmark the dataset, we propose TGIF, an efficient neural architecture that captures the dialog context, speaker dynamics, and emotion sequence in a conversation using Transformer encoders and stacked GRUs. Our evaluation yields state-of-the-art performance against five baselines adopted for our task. A detailed diagnosis of the competing models further illustrates the advantages and pitfalls of the neural architecture.

Code (0)

등록된 구현이 없습니다.

Methods 이 논문이 사용한 방법론

Similar Papers 제목 키워드 기반