By taking a closer look into the related repository folder for approaches, aimed at emotion causes extraction, I noticed it majorly exposes application of:
Recurrent NN approaches
Encoder-based Transformers (BERT) including spans predictions

However there is a gap in decoder based transformers, and application of instruction tuned models in particular

With that, and to fill this gap, I can recommend embedding the following and personally tested approach based on transformers and Flan-T5:

⭐https://github.com/nicolay-r/THOR-ECAC

That is a part of the following studies:
https://arxiv.org/abs/2404.03361

With the repository ⭐ above there are several Google colabs for quick launching