Multitask learning as a question answering

摘要

An approach to multitask learning as a question response isIncludes a method for trainingThe method isReceive multiple training samples including training samples from multiple task typesA training sample is presented to the neural model to generate an answer.About each training sample presentedDetermine the error between the generated answer and the natural language ground truth answerAndIncluding adjusting the neural model parameters based on the error.Each of the training samples includes the context of the natural language, question, and ground truth answer.The order in which the training sample is presented to the neural model includes first selecting a training sample according to a first training strategy and switching to selecting a training sample in accordance with a second training strategy.In some embodiments, the first training strategy is a sequential training strategy, and the second training strategy is an integrated training strategy.