METHOD FOR CONSTRUCTING SQL STATEMENT BASED ON ACTOR-CRITIC NETWORK

摘要

The invention discloses a method and a device for constructing a SQL statement based on reinforcement learning, wherein the method includes: initializing an actor-critic network parameter; acquiring a sequence pair of natural language and real SQL statement from a data set; inputting a natural language sequence into an actor network encoder, and inputting a real SQL sequence into a critic network encoder; using an encoded hidden state as an initialized hidden state of a corresponding decoder; gradually predicting, by an actor network decoder, a SQL statement action, and inputting the SQL statement action to a critic network decoder and an environment to obtain a corresponding reward; and using a gradient descent algorithm to update the network parameters, and obtaining a constructing model of the natural language to the SQL statement after repeated iteration training.