Joan Plepi

My current research activity is based primarily in contextual user embeddings based on flexible modeling of the user and network over conversational and social contexts using recurrent or attentive deep learning architectures. I am focused in both static and temporal analysis of user features in order to improve subjective NLP tasks. Currently, I am focusing on modeling different user’s perspective on social media. Moreover, I am working on modifying Transformer architectures (encoder-decoder like Flan-T5, or decoder-only like GPT, LLAMA) in order to include auxilary information for personalization.

Previously, I have worked in combining knowledge graphs and deep learning techniques in complex question answering domain.

Areas of Interest

- Fusing knowledge graph with LLMs
- Personalization using social networks for different NLP tasks
- Combining LLMs with GNNs
- Modifying LLMs to induce auxilary information for controllable generation
- Prompt/Prefix tuning of LLMs