Принудительное обучение

AbstractRecent approaches to question generation have used modifications to a Seq2Seq architecture inspired by advances in machine translation. Models are trained using teacher forcing to optimise only the one-step-ahead prediction. …...

AbstractMorphological inflection, like many sequence-to-sequence tasks, sees great performance from recurrent neural architectures when data is plentiful, but performance falls off sharply in lower-data settings. We investigate one aspect of …...

AbstractExposure bias and poor translation diversity are two common problems in neural machine trans-lation (NMT) which are caused by the general of the teacher forcing strategy for training inthe NMT …...

AbstractTeacher-forcing is widely used in training sequence generation models to improve sampling efficiency and to stabilize training. However, teacher-forcing is vulnerable to the exposure bias problem. Previous works have attempted …...