G2SERC: Unifying Graph Context Encoding and Sequential Emotion Decoding for Emotion Recognition in Conversation

ACL ARR 2026 January Submission3535 Authors

Keywords: Emotion Recognition in Conversations, Graph Neural Networks, Temporal Emotion Dynamics

Abstract:

Emotion Recognition in Conversation (ERC) aims to infer the emotions expressed in dialogues, where a relatively stable dialogue-level affective context coexists with transient utterance-level dynamics.
Prior work often emphasizes either sequential modeling of local context or graph-based aggregation of global dependencies, leaving the interaction between global atmosphere and evolving utterance emotions under-explored.
In this paper, we propose G2SERC, a graph-to-sequence framework that unifies relation-aware graph context encoding and sequential emotion decoding.
G2SERC first builds a speaker-aware heterogeneous graph over conversation and employs a relation-aware graph encoder to derive dialogue-level affective context and speaker-level affective priors.
A coupled recurrent decoder then tracks utterance dynamics while updating speaker-specific affective states, enabling emotion prediction conditioned on both dialogue evolution and speaker trajectories.
Extensive experiments show that G2SERC consistently outperforms strong baselines and achieves state-of-the-art performance.
Additional analyses demonstrate improved robustness to local emotional perturbations and substantiate the benefit of integrating global and speaker-aware signals.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 3535