Relational Dynamics of Emotion Recognition in Conversation

This direction is merged into another direction: context-aware emotion recognition (dialogue history integration). Its last version stays readable and citable here.

Version 1 of 129 papers30 conceptsfrozen 2026-10-04

In isolated sentiment analysis, a sentence like "That’s great" is almost always classified as positive. But in a conversation, its meaning is volatile. If said after a friend shares good news, it is joyful; if said after a sarcastic remark about a broken car, it is frustrated.

Emotion Recognition in Conversation (ERC) shifts the focus from static text to the moving parts of human interaction. To understand what someone feels, a model cannot just look at the words; it must track the "emotional inertia" of the speaker and the "inter-speaker influence" of their partners.

The Two Forces of Conversational Emotion

Human emotion in dialogue is governed by two primary forces. The first is emotional inertia: our tendency to stay in a specific emotional state over time. If you were angry two minutes ago, you are likely still frustrated now. The second is inter-speaker influence: the way we react to the emotional cues of others.

Early models often struggled to balance these. Devamanyu Hazarika, Soujanya Poria, and colleagues addressed this with the Conversational Memory Network (CMN). By using a specialized memory architecture, CMN explicitly models how a speaker’s past feelings persist while simultaneously accounting for how they are nudged by their interlocutor. This dual-tracking is essential because an utterance's emotional state is often contingent on the preceding history, as noted by Chunyan Miao and Peixiang Zhong in their work on RCNN-based classification.

From Sequences to Relational Graphs

While traditional Recurrent Neural Networks (RNNs) process dialogue as a linear sequence, they often suffer from "forgetting" long-term context. In a multi-party conversation, the most relevant context for Speaker A might not be the immediately preceding sentence from Speaker B, but rather something Speaker A said three turns ago.

To solve this, Deepanway Ghosal and Navonil Majumder proposed DialogueGCN, which treats a conversation as a directed graph. In this graph, nodes represent utterances, and edges represent the relationships between them—whether they come from the same speaker or a different one. This allows the model to "jump" across the sequence to find the most influential context.

This graph-based intuition was further refined by Abhinav Joshi and Ashutosh Modi in COGMEN, which combines the global perspective of Transformers with the local relational modeling of Graph Neural Networks (GNNs). By looking at both the broad conversation flow and specific speaker-to-speaker dependencies, these models capture the "multi-directional dependencies" that define real-world social dynamics.

Identifying the Instigators of Change

A major challenge in ERC is the "emotion flip"—a sudden shift from one state to another. Juan Yang and Puling Wei argue in their work on GAT-CRESA that models must be "emotion-shift aware." If a speaker suddenly becomes angry, the model needs to reason about why that happened.

This led to the concept of Emotion Flip Reasoning (EFR). Shivani Kumar and Md Shad Akhtar introduced this task to identify the psychological "instigators"—such as threat or annoyance—that trigger these shifts (Emotion Flip Reasoning in Multiparty Conversations). Understanding the cause of an emotion (the cognition) is just as important as identifying the emotion itself (the affection), a concept explored by Jinfeng Zhou and Minlie Huang in the CASE model.

Empathy as a Collaborative Act

Finally, recent research suggests that emotion in dialogue is not just something a speaker has, but something speakers create together. Zhichao Xu and Jiepu Jiang posit that empathy is a collaborative phenomenon (Multi-dimensional Evaluation of Empathetic Dialog Responses). They found that the meaning of an empathetic response is often shaped by the subsequent reaction of the listener.

This relational view explains why analyzing individual utterances in isolation is insufficient. Whether it is recognizing humor in a Hindi sitcom (Dushyant Singh Chauhan, M2H2) or detecting frustration in a task-oriented helpdesk call (Shutong Feng, ERToD), the intuition remains the same: emotion is a product of the interaction between people, their shared history, and their immediate goals.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_relational_dynamics_of_emotion_recogniti_124392e5,
  title = {Relational Dynamics of Emotion Recognition in Conversation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relational-dynamics-of-emotion-recognition-in-conversation/rdr_124392e5b56d46b885b83f58a5cd438f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_relational_dynamics_of_emotion_recogniti_124392e5_v1,
  title = {Relational Dynamics of Emotion Recognition in Conversation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relational-dynamics-of-emotion-recognition-in-conversation/rdr_124392e5b56d46b885b83f58a5cd438f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}