Gerald is a Retreival and Generative Ensemble for Conversational AI made with BERT Transformer and the HuggingFace T5 Conditional Generator
Talk to Gerald Here!
What is Gerald?
An open-source and open-domain conversation AI model named GERALD (named after Gerald Salton) that uses a sequential ensemble of a retrieval-based and generation-based system to intelligently respond to user queries while maintaining contextual and structural relevance. The retrieval-based system has been trained on a large dataset based on conversational AI question-answer pairs and natural language conversations from multiple sources. The top k queries in the dataset that match closely with a user query are extracted and processed. These queries are then evaluated and passed on to the generative model, which intelligently generates new responses which are contextually relevant and structurally aligned with the user query. The main objective of the program is to generate queries intelligently.
The retrieval-based system uses a BERT transformer to contextually evaluate a user query against the dataset and the generative model uses a T5 transformer model which conditions on the retrieved text and the user query to generate new sequences. This ensemble system is known to outperform retrieval based and generation-based models working independently.