Multi-task learning as a question answering

摘要

The multi-task learning approach as a question response includes an input layer for encoding contexts and questions; a self-attention-based transformer that includes an encoder and decoder; and a first bidirectional for further encoding the encoder output. Long / short-term memory (biLSTM); long-short-term memory (LSTM) for generating context-adjusted hidden state from encoder output and hidden state; first based on first biLSTM output and LSTM output Includes an attention network for generating attention weights; a vocabulary layer that produces a vocabulary distribution state; a context layer that produces a context distribution state; a switch; Generate weights between the distribution states of, generate a compound distribution state based on the weights, and use the compound distribution state to select the answer word.