| Résumé | We introduce EmoAtlas, a computational library/framework extracting emotions and syntactic/semantic word associations from texts. EmoAtlas combines interpretable artificial intelligence (AI) for syntactic parsing in 18 languages and psychologically validated lexicons for detecting the eight emotions in Plutchik’s theory. We show that EmoAtlas can match or surpass transformer-based natural language processing techniques, BERT or large language models like ChatGPT 3.5 or LLaMAntino, in detecting emotions from Italian and English online posts and news articles (e.g., achieving 85.6% accuracy in detecting anger in posts vs the 68.8% value of ChatGPT and 89.9% value for BERT). EmoAtlas presents important advantages in terms of speed and absence of fine-tuning, e.g., it runs 12x faster than BERT on the same data. Testing EmoAtlas’ and easily trainable transformers’ relevance in a psychometric task like reproducing human creativity ratings for 1071 short texts, we find that EmoAtlas and BERT obtain equivalent predictive power (fourfold cross-validation, ρ ≈ 0.495, p < 10⁻⁴). Combining BERT’s semantic features with EmoAtlas’ emotional/syntactic networks of words gets substantially better at estimating creativity rates of stories (ρ = 0.628, p < 10⁻⁴). This indicates an interplay between the creativity of narratives and their semantic, emotional, and syntactic structure. Via interpretable emotional profiles and syntactic networks, EmoAtlas can also quantify how emotions are channeled through specific words in texts, e.g., how did customers frame their ideas and emotions towards ”beds” in hotel reviews? We release EmoAtlas as a standalone ”text as data” computational tool and discuss its impact in extracting interpretable and reproducible insights from texts. |
|---|
