When humans engage in conversation or follow a narrative, we rarely treat the present moment as an isolated event. If someone becomes suddenly angry, enthusiastic, or sarcastic, the trigger is frequently not the sentence that was just uttered. Instead, the true cause often lies in an event, a revelation, or a subtle shift in tone that occurred minutes or even hours prior.
However, standard natural language processing and multimodal models suffer from a profound proximity bias (or position bias). As noted by Fanfan Wang et al. in SemEval-2024 Task 3: Multimodal Emotion Cause Analysis in Conversations, conversational models heavily rely on the heuristic that the immediate predecessor in a dialogue is the cause of the current emotional state. This local focus creates a major bottleneck when trying to understand the true causal dynamics of human interaction.
Why Local Context Fails
Why do sequential models struggle with long-range dependencies in dialogue? The core issue is that the chronological flow of a conversation does not match its logical or emotional architecture.
In multi-turn or multi-party dialogues, information is highly distributed and non-linear. For instance, Ashutosh Modi et al. demonstrate in Multi-Task Learning Framework for Extracting Emotion Cause Span and Entailment in Conversations that emotional triggers are frequently located three or more utterances back. When a model restricts its view to a narrow, fixed-length context window, it completely misses these distal causes. As Jianwu Dang 0001 argues in Dimensional Emotion Prediction Based on Interactive Context in Conversation, fixed-length windows fail to capture the cumulative influence of historical interactions, ignoring the slow-varying emotional trajectories that develop over time.
This problem is magnified in narrative media like movies. Analyzing a single video frame or scene in isolation fails to capture what Emmanuel Dellandrea et al. describe as the "momentum" of preceding scenes in The MediaEval 2016 Emotional Impact of Movies Task. A jump scare is only frightening because of the preceding silence; a character's current disposition, as Dhruv Srivastava et al. point out in How You Feelin'? Learning Emotions and Mental States in Movie Scenes, is the cumulative result of an entire narrative arc rather than the immediate visual stimuli.
Structural Solutions to Bridge the Gap
To overcome proximity bias, researchers have shifted from treating dialogue as a simple linear sequence to modeling it as a structured, multi-layered network.
Discourse and Ancestry Graphs
Instead of forcing a model to look strictly backward in time, graph-based architectures map the explicit logical relationships between turns. For example, Henghui Zhu et al. propose a Masked Hierarchical Transformer in Who Did They Respond to? Conversation Structure Modeling Using Masked Hierarchical Transformer that allows utterances to attend specifically to their conversational ancestors rather than just the chronologically prior message. Similarly, Fandong Meng et al. utilize social commonsense knowledge graphs in Neutral Utterances are Also Causes: Enhancing Conversational Causal Emotion Entailment with Social Commonsense Knowledge to bridge the reasoning gap between distant, seemingly neutral statements and their eventual emotional effects.
Hierarchical State Tracking
Another powerful intuition is separating local token-level dynamics from global conversation-level dynamics. Models like DialogueRNN, developed by Navonil Majumder et al. in DialogueRNN: An Attentive RNN for Emotion Detection in Conversations, explicitly track individual speaker identities and global context over dozens of turns. This allows the network to maintain an internal representation of a speaker's emotional "inertia," making it easier to link a current reaction to a trigger that occurred far back in the history. Shanglin Lei addresses this in Watch the Speakers: A Hybrid Continuous Attribution Network for Emotion Recognition in Conversation With Emotion Disentanglement by combining recurrent architectures with attention-based attribution mechanisms, balancing local temporal continuity with long-range dependencies.
Causal Discovery and Filtering
Simply expanding the context window often introduces noise and spurious correlations, as models might latch onto irrelevant keywords from earlier turns. To combat this, Lizhen Qu et al. introduce a method based on causal discovery in Less is More: Mitigate Spurious Correlations for Open-Domain Dialogue Response Generation Models by Causal Discovery. By mathematically isolating the specific past utterances that causally drive a response, the framework filters out the background "chitchat," allowing the system to maintain long-term coherence without getting distracted by non-proximal noise.
Key Foundational Work
Context-sensitive multimodal emotion recognition from speech and facial expression using bidirectional LSTM modeling (2010) — contributed 1 concept elements to this direction.
Recursive Neural Networks for Coding Therapist and Patient Behavior in Motivational Interviewing (2015) — contributed 1 concept elements to this direction.
Towards Conversation Entailment: An Empirical Investigation (2010) — contributed 1 concept elements to this direction.