Modelling a Conversational Agent with Complex Emotional Intelligence

Billal Belainine, Fatiha Sadat, Hakim LounisAAAI 2020 · DOI

Abstract

This paper proposes a novel architecture for conversational agents designed to move beyond basic emotional labels toward a nuanced handling of complex, multi-layered human feelings. By mapping emotions into a four-dimensional vector space based on Plutchik’s wheel of emotions, the researchers aim to create chatbots capable of emotionally consistent, multi-turn conversations that reflect an internal "personality."

Introduction

While modern conversational agents are increasingly proficient at task-oriented dialogue, they often lack true emotional intelligence. Most existing models, such as the Emotional Chatting Machine, focus on a limited set of discrete basic emotions like joy, anger, or sadness. However, human experience is characterized by complex "dyads"—combinations of basic emotions like optimism (joy + anticipation) or remorse (sadness + disgust). Furthermore, many current models treat dialogue turns in isolation, failing to maintain an emotional memory that evolves over time. This research seeks to bridge this gap by designing an agent that can perceive, process, and express these complex emotional states over extended interactions.

Proposed Architecture

The proposed framework is built upon a two-unit architecture designed to handle both the perception and expression of emotion:

  1. Natural Language Understanding (NLU) Unit: This unit is responsible for linguistic and emotional processing. It decomposes sentences into morphosyntactic forms and generates semantic vectors via word embeddings. It then detects basic emotions to predict the agent's "emotional reflex"—its internal state.
  2. Natural Language Generation (NLG) Unit: This unit recomposes words into syntactically and semantically valid responses that reflect the calculated emotional state, ensuring the output is contextually and emotionally appropriate.
Figure 1: The proposed architecture for the conversational agent, highlighting the flow from NLU to NLG.

Emotional Representation and Mapping

To represent complex emotions mathematically, the authors utilize Robert Plutchik’s wheel of emotions. Every emotional state is mapped as a vector in a 4-dimensional space with values ranging from -1 to 1. Each dimension represents a pair of contradictory basic emotions (e.g., Joy vs. Sadness, Trust vs. Disgust).

Complex emotions, or "advanced emotions," are defined as combinations of these dimensions using a transition matrix. For instance, Optimism is represented by a positive value (1) in both the "Anticipation-Surprise" and "Joy-Sadness" dimensions, while Remorse (listed as Remord in the source) involves negative values in the "Joy-Sadness" and "Sympathy-Disgust" dimensions.

Figure 2: Mapping of advanced emotions (primary dyads) into the 4D vector space.

Key Insights

  • Emotion as a Coordinate System: Treating emotions as vectors allows the model to mathematically "blend" feelings, enabling smooth transitions between states rather than abrupt shifts.
  • The Emotional Reflex: The agent predicts its own internal reaction to user input based on its "personality," rather than simply mirroring the user's tone.
  • Temporal Interference: Emotions detected in early stages of a conversation influence later stages. This creates a sense of emotional memory, where the mood of the interaction builds and evolves over time.

Current Progress and Future Work

As this work was presented as a doctoral project, the results are currently focused on architectural logic and mapping. The authors have successfully mapped primary dyads into a 3-valued label system (-1, 0, 1) across the four dimensions.

Future work will focus on the final evaluation phase, comparing the 4D vector approach against standard cross-entropy models using datasets like DailyDialog and MojiTalk. Key challenges remaining include managing the computational complexity of 4D vector spaces in long-term memory and refining the agent's ability to "steer" the emotional trajectory of a conversation.

Summary updated

BibTeX
@inproceedings{belainine2020modelling,
  author = {Billal Belainine and Fatiha Sadat and Hakim Lounis},
  title = {{Modelling a Conversational Agent with Complex Emotional Intelligence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7127},
}
Figure