An information fusion based approach to context-based fine-tuning of GPT models

In the new era of Artificial Intelligence (AI), Generative Pre-Trained Transformer (GPT) has emerged as a central technique for generating human-like texts. Over recent years, there has been a growing trend towards using GPT for building chatbot systems. However, pre-trained GPT models lack context awareness which can result in awkward dialogue in specific contexts. In this research, we propose a new information fusion based approach to fine tuning GPT models based on contextual data and two scenarios of evidence combination by means of Dempster–Shafer theory of evidence. To this end, we first design a Transformers-based dialog classification model to be trained with the contextual data, which is then used jointly with additional pre-trained models as sources of evidence for judging the output of a GPT model as a context-appropriate response. Two scenarios for modeling and combining evidence provided by the context-based dialog classification model and pre-trained models are also proposed. We conduct a set of experiments on several datasets associated with specific contexts to demonstrate the effectiveness of the proposed approach. The empirical results show that it can improve the contextuality of general GPT-2 and GPT-3.5 models in most cases of the testing datasets.