Memory network-based intention recognition method under multi-round dialogues

Abstract

The invention discloses a memory network-based intention recognition method under multi-round dialogues. A neural network model combining a recurrent neural network, a memory network and an attentionmechanism is mainly adopted. The method comprises the steps of firstly performing mathematical processing on a language; secondly encoding each round of the dialogue by utilizing the recurrent neuralnetwork to obtain an encoding vector of the dialogue; thirdly storing historical dialogue information by utilizing an external memory; fourthly, selecting the historical dialogue information related to an intention of the current round of the dialogue from the memory through an attention mechanism, thereby obtaining an encoding vector of the historical information; fifthly, for the current dialogue encoding vector and the historical dialogue encoding vector, judging whether the historical information is introduced in a classifier or not by utilizing a control gate, thereby obtaining encoding information which is finally used for classification; and finally, obtaining the intention of each round of the dialogue by utilizing the multi-label classifier.

Description

technical field [0001] The invention belongs to the dialogue field of natural language processing, and relates to a neural network learning method of a memory network. Background technique [0002] With the continuous improvement of people's living standards, the emergence of artificial intelligence equipment actually meets the "small rigid needs" of the public's psychology, such as smart speakers, one of the artificial intelligence equipment. Smart services such as playing music through voice control, or turning off the lights while lying in bed in winter, have greatly facilitated people's lives to a certain extent. The explosive growth of the domestic smart speaker market in 2017 is astonishing, and the sales data in 2017 is enough to show the popularity of smart speakers in China. This year was the fastest-growing year for the smart speaker market, and many brands quickly won the market through their own software or hardware advantages. Of course, behind the success, th...

AI Technical Summary

Problems solved by technology

The memory of the cyclic neural network is an internal memory method, which is realized by rnnCell or lstmCell, but the memory capabilities of rnn and lstm are really limited, and at most a dozen time steps can be memorized

Images