Emotion Elicitation Under Audiovisual Stimuli Reception: Should Artificial Intelligence Consider the Gender...

Current automatic safety systems often lack the precision required to identify high-stress emotional states like panic or fear effectively. No prior work has fully resolved how diverse stimuli sources influence the accuracy of these detection models. Researchers frequently struggle to obtain datasets that provide both discrete and quantitative emotional labels. That uncertainty drove the need for a standardized collection of audiovisual clips. Prior research has shown that emotional responses can vary significantly across different demographic groups. However, existing models rarely account for these variations when processing sensitive scenarios. This gap motivated the development of a specialized resource for training machine learning algorithms. The authors address these limitations by creating a comprehensive database focused on eliciting a wide range of human affective states.

The primary aim of this research is to evaluate whether artificial intelligence systems should incorporate a gender perspective when processing emotional responses to audiovisual stimuli. The authors seek to improve the efficiency of automatic safety solutions designed to protect vulnerable populations. They address the challenge of discerning risky situations, such as those involving panic or fear, with greater accuracy. The team identifies a need for high-quality datasets that allow for both discrete and quantitative emotional categorization. By creating the UC3M4Safety database, they provide a resource tailored for training machine learning algorithms. The study investigates how different sources of stimuli influence the elicitation of human affective states. They specifically focus on scenarios like gender-based violence to determine if men and women react differently. This investigation motivates the development of more precise tools that account for demographic variations in emotional processing.

The team constructed the UC3M4Safety database to provide a standardized set of clips for affective computing research. They recruited 1332 participants to view these materials and report their internal states. Each video received an average of 84 evaluations to ensure high reliability. The investigators applied both discrete and quantitative labeling techniques to categorize the reported feelings. They utilized the Pleasure-Arousal-Dominance Affective space to map these responses numerically. Statistical comparisons between male and female subjects were performed to identify potential discrepancies in reaction patterns. The researchers processed all gathered information using RStudio to maintain rigorous analytical standards. This approach ensured that the resulting dataset could effectively train machine learning models for detecting panic or fear.

The study successfully generated a comprehensive database of audiovisual clips that consistently trigger a wide range of human emotions. Participants demonstrated a high level of agreement regarding the emotional content of the selected stimuli. The researchers identified distinct emotional responses when comparing the 811 women and 521 men involved in the trials. These gender-based differences were particularly evident in scenarios depicting gender-based violence. The dataset provides both discrete categories and quantitative values within the Pleasure-Arousal-Dominance Affective space. Each video achieved an average of 84 responses, with a standard deviation of 22. These metrics confirm the suitability of the database for training various machine learning algorithms. The findings indicate that incorporating these gender-specific patterns significantly enhances the precision of automatic systems designed to detect risky situations.

The authors propose that their new audiovisual database provides the necessary foundation for training more precise safety-oriented artificial intelligence systems. They suggest that incorporating gender-specific response patterns improves the ability of algorithms to detect fear or panic. The researchers indicate that their findings support the development of tools aimed at protecting vulnerable populations. They maintain that the observed differences in emotional reactions between men and women are significant enough to warrant inclusion in future model designs. The study demonstrates that discrete and quantitative emotional categorization is achievable through standardized stimulus selection. They conclude that their dataset is well-suited for machine learning applications requiring high levels of participant agreement. The team emphasizes that focusing on scenarios like gender-based violence enhances the relevance of automated detection solutions. They suggest that future implementations should prioritize these nuanced emotional profiles to ensure greater system reliability.