Artificial Empathy Classification: A Survey of Deep Learning Techniques, Datasets, and Evaluation Scales

Sharjeel Tahir, Syed Afaq Ali Shah, Jumana Abu-KhalafarXiv 2023 · DOI · arXiv

Abstract

Over the last decade, researchers in machine learning and assistive developmental robotics have pivoted toward Artificial Empathy (AE) as a critical frontier for human-robot interaction (HRI). While humans naturally develop empathy from birth, instilling this capability in machines remains a formidable challenge. This paper surveys the current landscape of AE, examining the deep learning architectures, datasets, and evaluation scales used to bridge the emotional gap. By analyzing the standard workflow—from emotion recognition to response generation—we highlight state-of-the-art methods and identify the urgent need for standardized evaluation metrics.

The Challenge of Artificial Empathy

The primary hurdle in AE research is the transition from simple pattern matching to genuine emotional resonance. The field currently struggles with "distorted realities," where rule-based systems are often marketed as empathetic, and researchers may anthropomorphize AI that merely identifies patterns without true understanding.

Figure 1: The hierarchy of empathy, moving from basic emotional contagion to complex compassionate action.

As illustrated in the hierarchy above, empathy is not a monolithic trait but a tiered capability. Most current systems operate at the level of emotional or cognitive empathy, while compassionate empathy—the ability to take appropriate action based on perceived distress—remains the ultimate goal for assistive robotics.

Methodological Framework

The standard AE workflow consists of three distinct stages: Emotion Recognition (ER) using retrieved features, Analysis of the perceived emotion to choose an action, and the execution of a Response Action.

Deep Learning (DL) has become the engine of this workflow. Transformers like BERT and GPT-3 dominate textual AE, while CNNs (e.g., VGGFace) and LSTMs handle multi-modal inputs. Furthermore, Reinforcement Learning (RL) is increasingly used to optimize empathic policies, distinguishing between Extrinsic Motivation (aiming for user state change) and Intrinsic Motivation (prioritizing conversational quality and human needs as the primary reward).

The Data Landscape and Evaluation Gap

A significant imbalance exists in the available training data. Textual datasets like Facebook’s Empathic Dialogues (FED) are abundant, making up over half of the research focus. In contrast, naturalistic visual and audio datasets are scarce, often relying on scripted or controlled environments.

Perhaps the most critical issue identified is the "BLEU Fallacy." Standard NLP metrics like BLEU or ROUGE measure grammatical similarity, which does not necessarily correlate with emotional resonance. A machine's response can be grammatically perfect yet emotionally tone-deaf.

Figure 2: Comparison of evaluation metrics across different model configurations, highlighting the nuances in measuring empathy, relevance, and fluency.

As shown in the table above, different model configurations yield varying results across empathy and fluency scales. While human evaluation remains the gold standard (used in 65% of research), the field lacks a universally accepted, autonomous metric that accurately captures the "empathetic" quality of a response.

Future Directions

To move AE from "monotonous and controlled" to naturalistic and trustworthy, the research community must focus on:

  • Standardization: Establishing a generalized definition of AE to ensure consistent measurement.
  • Non-Verbal Cues: Moving beyond facial expressions to include tone of voice and body gestures.
  • Explainability (XAI): Understanding the "why" behind a robot's empathetic choice to build user trust.

Developing empathy-specific metrics that correlate with human perception is essential for the next generation of truly assistive AI.

Summary updated

Abstract

From the last decade, researchers in the field of machine learning (ML) and assistive developmental robotics (ADR) have taken an interest in artificial empathy (AE) as a possible future paradigm for human-robot interaction (HRI). Humans learn empathy since birth, therefore, it is challenging to instill this sense in robots and intelligent machines. Nevertheless, by training over a vast amount of data and time, imitating empathy, to a certain extent, can be possible for robots. Training techniques for AE, along with findings from the field of empathetic AI research, are ever-evolving. The standard workflow for artificial empathy consists of three stages: 1) Emotion Recognition (ER) using the retrieved features from video or textual data, 2) analyzing the perceived emotion or degree of empathy to choose the best course of action, and 3) carrying out a response action. Recent studies that show AE being used with virtual agents or robots often include Deep Learning (DL) techniques. For instance, models like VGGFace are used to conduct ER. Semi-supervised models like Autoencoders generate the corresponding emotional states and behavioral responses. However, there has not been any study that presents an independent approach for evaluating AE, or the degree to which a reaction was empathetic. This paper aims to investigate and evaluate existing works for measuring and evaluating empathy, as well as the datasets that have been collected and used so far. Our goal is to highlight and facilitate the use of state-of-the-art methods in the area of AE by comparing their performance. This will aid researchers in the area of AE in selecting their approaches with precision.
BibTeX
@misc{tahir2023artificial,
  author = {Sharjeel Tahir and Syed Afaq Ali Shah and Jumana Abu-Khalaf},
  title = {{Artificial Empathy Classification: A Survey of Deep Learning Techniques, Datasets, and Evaluation Scales}},
  year = {2023},
  eprint = {2310.00010},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.00010},
}
Figure