A method for knowledge construction from natural language based on reinforcement learning

It is promising to raise the intelligence level of service robot by extracting information from natural language. The problem has been noticed that NLP system frequently mislabel passive voice verb phrases as being in the active voice on the condition without auxiliary verb, which will reduce the correctness of produced knowledge based on natural language. A method based on reinforcement learning for household knowledge construction is proposed in this paper, according to the problem. Firstly, a dynamic simulation platform modelled on the laboratoryl is built. Secondly, a system of justification is designed as agent to produce scores of the action sequences from natural language. Thirdly, with the combination of above parts, we adopt reinforcement learning algorithm for training to obtain the appropriate compositions of action sequences. As preprocessing has been made to reduce the complexity of the actions and states, expected results can be obtained with this method.