Complex Emotional Intelligence Learning Using Deep Neural Networks (Student Abstract)

Belainine Billal, Fatiha Sadat, Hakim LounisAAAI 2020 · DOI

Abstract

Traditional sentiment analysis often oversimplifies human emotion into binary or basic categories. This research introduces a framework for detecting complex emotional states—such as Love, Shame, and Guilt—by leveraging Plutchik’s wheel of emotions. By employing transfer learning from the DeepMoji model and a novel 4-dimensional vector space representation, the study demonstrates how basic emotions can be mathematically modeled as components of sophisticated human experiences.

Introduction

Human psychology is characterized by a rich tapestry of emotions that are rarely "atomic." While most emotion recognition systems focus on basic states like joy or sadness, real-world interactions involve "complex emotions" formed by blending these basics. For instance, psychological theory suggests that Love is a combination of Joy and Trust, whereas Shame arises from Fear and Disgust. This paper addresses the challenges of data scarcity and modeling complexity in identifying these nuanced states.

Methodology: The 4D Vector Space

The core of the proposed approach is grounded in Plutchik’s Theory of Emotions, which organizes eight basic emotions into four pairs of opposites. The researchers defined a base ($B$) consisting of these four axes: (Joy, Sadness), (Trust, Disgust), (Fear, Anger), and (Surprise, Anticipation).

To represent these mathematically, the system utilizes a Sequence-to-Vector (Seq2vec) model to assign values (1, -1, or 0) across these four dimensions. A key innovation is the use of a Transition Matrix ($W$) to map basic emotion vectors to "Primary Dyads" (complex emotions). This allows the model to decompose complex labels found in existing datasets into their constituent basic components for training, and then reconstruct them during inference.

Table 1: Mapping of advanced emotions (primary dyads) to the four primary emotional axes.

To overcome the lack of labeled data for complex emotions, the researchers employed transfer learning using DeepMoji, a model pre-trained on millions of emoji-rich tweets. This provided a robust baseline for emotional understanding that was then fine-tuned for multi-label classification.

Results and Discussion

The model's performance was evaluated on its ability to detect both basic emotional axes and complex dyads. On the basic axes, the model achieved F1-scores ranging from 0.49 for Joy/Sadness to 0.58 for Anger/Fear. Crucially, the proposed model demonstrated superior or competitive performance compared to the standard DeepMoji baseline when identifying complex emotional states.

Table 2: Performance comparison between the proposed model and DeepMoji baselines across complex emotions.

As shown in the results, the model achieved an average F1-score of 0.58 for Love, 0.54 for Shame, and 0.54 for Guilt. Notably, the model excelled in "Exact Match" scenarios, indicating a high proficiency in correctly identifying both underlying components of a complex emotion simultaneously.

Conclusion and Future Work

By grounding deep learning in established psychological frameworks, this research provides a scalable method for interpreting complex human emotions from text. While challenges remain—such as data imbalance for emotions like "disgust" and "anticipation"—the use of a 4D vector space offers a promising path forward. Future research will focus on refining the transformer decoder architecture and expanding the datasets to cover the full spectrum of Plutchik’s emotional wheel.

Summary updated

BibTeX
@inproceedings{billal2020complex,
  author = {Belainine Billal and Fatiha Sadat and Hakim Lounis},
  title = {{Complex Emotional Intelligence Learning Using Deep Neural Networks (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7149},
}
Figure