METHOD AND SYSTEM FOR TRAINING NEURAL SEQUENCE-TO-SEQUENCE MODELS BY INCORPORATING GLOBAL FEATURES

摘要

These are methods for training a neural sequence-to-sequence (seq2seq). The processor receives training data including a model and a plurality of training source sequences and corresponding training target sequences, and generates corresponding predicted target sequences. The model parameters are local loss in the predicted target sequences based on the comparison of the predicted target sequences to the training target sequences, and between the predicted target sequences and the training target sequences given the training source sequences. Is updated to reduce or minimize both the expected loss of one or more global or semantic features or constraints of. Expected loss is based on global or semantic features or constraints of general target sequences given the general source sequences.