Enhancing LLM Conversational Acuity Using Pragmatic Measures
Andrew Han, Nishanth Koushik, Nachiket Bidarkundi, Michael Naeim Shehata, Vibha Kunchipudi, Tunar Mammadli, Shourya Mehta, Osher Lerner · 2024
AI-driven communication has the potential to transform society with superhuman capabilities such as real-time multilingual translation, predictive text generation, and personalized content creation. However, current Large Language Models (LLMs) like OpenAI's GPT-3.5 Turbo and 40-mini often struggle to capture the nuanced writing styles, tones, and behavioral characteristics of individual users. While fine-tuning is a common approach, existing techniques focus primarily on task-specific performance and tend to neglect the integration of personal conversational elements and non-verbal cues. As a result, these methods often fail to preserve the unique conversational tone and vocabulary of individual users. Our work addresses these limitations by analyzing similarity, linguistic, and psychological metrics to fine-tune OpenAI's GPT-3.5 Turbo and 4o mini, improving contextual accuracy, and generating responses that better align with conversational tone and user-specific vocabulary.