Joint emotion label space modeling for affect lexica

Emotion lexica are commonly used resources to combat data poverty in automatic emotion detection. However, vocabulary coverage issues, differences construction method and discrepancies framework representation result a heterogeneous landscape of detection resources, calling for unified approach utilizing them. To this, we present an extended lexicon 30,273 unique entries, which is merging eight existing by means multi-view variational autoencoder (VAE). We showed that VAE valid combining with different label spaces into joint space chosen number dimensions, these dimensions still interpretable. tested the utility employing values as features model. found outperformed individual lexica, but contrary our expectations, it did not outperform naive concatenation although contribute when added extra lexicon. Furthermore, using information additional on top state-of-the-art language models usually resulted better performance than no was used.