Autonomous Generation Mechanism Of Service Knowledge For Household Robots Based On Natural Language

With complexity and diversity,home services can be regarded as a logical combination of simple actions,satisfying people's requirements under the cooperation of subtasks.During the process of performing home services,robots should consider their own environment and are required to possess the conception for service execution,also called service knowledge.Under the supervision of researchers,traditional service knowledge is built oriented to the entire task of home service,by programming the logic of subtasks with service rules.Knowledge construction with this method costs manual efforts,and the corresponding knowledge lacks universality.There exist service elements,including service tools,actions,logical relationships between subtasks,in natural language information related to home services,which can be used to provide strategic guidance for robots to perform services.Based on this point,it is valuable to make research on how robots learn service-related natural language information and translate them into service knowledge,which helps to increase the intelligence and service techniques of robots.In view of the complex home environment,this dissertation deeply studies key technologies of Natural Language Understanding and methods of knowledge construction,and is disserted on four aspects:information representation,strategy generation,service elements extraction and knowledge construction.The main contents in this dissertation are as follows:1.In order to represent natural language information related to home services properly,a document representation model based on tool information is proposed.It can increase the accuracy of information representation and make the document representation model give more attention on service-related vocabulary.To begin with,an end-to-end model,tool cognition layer,is constructed to produce tool information corresponding to service requirements.Then,tool information is introduced for emphasizing the importance of tool information during the training process of word vectors.Finally,tool information based attention mechanism is built to associate tool information to sentence vectors,and the document representation relevant to home services is completed.2.Effective service strategies can provide guidance on how to perform home services.Aiming at the generation of effective strategies,a reinforcement learning method with tool information and lexical relevance as the guide is proposed to produce effective service strategies.In the aspect of accuracy in strategy generation,reward functions based on tool information are designed,to guide the generation of effective strategies.In the aspect of stability in strategy generation,lexical relevance in introduced to enhance the model stability.To address the scarcity of training data in the field of home service,a strategy generation method based on adversarial learning is proposed,to simulate the real distribution of strategies and generate service strategies with reference value.3.Aiming at the extraction of service elements in strategies,a reinforcement learning method for action sequence generation is proposed,to realize the transformation from strategy information to action sequences.Service elements in action sequences are conducive to the autonomous generation of follow-up service knowledge.Firstly,semantic parsing technology is adopted to obtain service elements contained in service strategies.Secondly,a service evaluation method based on object states is proposed to evaluate the validity of action sequences.Then,a simulation platform oriented to home services is built to simulate the service execution of robots.Finally,the action sequence generation with reinforcement learning is completed by using simulation platform as interactive environment and combining service evaluation method.4.For the construction of service knowledge,a method for building a hierarchical object knowledge base is proposed to endow action sequences with object operation knowledge and constitute service knowledge.Aiming at the quantitative modification of service knowledge,a method of knowledge modification based on user evaluation is proposed.By analyzing user evaluation information,relevant service actions can be adjusted and service knowledge modification is completed.Firstly.user evaluation is extended with service knowledge information,in order to increase the reasoning ability of the revision model.Then,revision rules are designed with Semantic Web Rule Language to mapping revision conditions to conclusions.Finally,a revision model with attention mechanism is constructed to modify service knowledge according to user evaluation.