Sapientino is a BERT-based web application that solves an open domain QA task ( competition launched by Google in 2019), namely it provides answers for questions expressed in naturla language.
Any user can ask questions to Sapientino, which will look for pertinent Wikipedia pages and query our BERT-based model for answering such questions. This API is easily transferable to new implementations, e.g. a vocal assistant.
Let us dig into some technicalities: Natural Questions (the dataset on which Sapientino was fine-tuned) allows three kinds of answers: the “simple” yes/no answers, the longer and more articulated long answers, and the natural trade-off between expressive power and succinctness, the “so-called” short answers, which aim to enclose the answer in a single and possibly short sentence.
The fine-tuning process was held working on both ALBERT and BERT models, performing a grid search for hyper-parameters selection. As a result, we observed that the BERT model has the best performances.