Pre-training on Multi-modal for Improved Persona Detection Using Multi Datasets

Abstract

AbstractPersona identification helps AI-based communication systems provide personalized and situationally informed interactions. This paper introduces pre-training on CNN, BERT, and GPT models to improve persona detection on PMPC and ROCStories datasets. Two speakers with different personalities have dialogues in the PMPC dataset. The challenge is to match each speaker to their persona. The ROCStories dataset contains fictional character traits and activities. Our study uses transformer-based design to improve persona detection using ROCStories dataset external context. We compare our method to leading models in the field. We found that pre-training and fine-tuning on several datasets improves model performance. External context from tale collections may improve persona detection algorithms and help understand human personality and behavior. Our study found that pre-training CNN, BERT, and GPT models improves persona detection, improving user experiences and communication. The method could be used in chatbots, personalized recommendation systems, and customer support. Additionally, it can help create AI-driven communication systems with tailored, context-aware, and human-like interactions.