-- LEARNING METHOD FOR REPEATING GENERATION CONTROL OF SEQUENCE TO SEQUENCE MODEL

摘要

Disclosed is a method of learning a sequence-to-sequence summary model. The method includes: an encoder outputting a hidden state vector for a summary generated by the sequence-to-sequence summary model, and a decoder, for each step using the hidden state vector, each word in the summary Calculating the probability to be generated, generating a word corresponding to the highest probability among the calculated probability, and using the history of words generated by the decoder, the sequence-to-sequence summary model previously output And learning not to duplicate words.