Research On Human-Robot Dialogue Parsing For Home Service Robot

The ability for home service robot of parsing instructions determines whether the human-robot interaction process is unobstructed,which is related to whether home service robot can play an essential role in users’ daily life.The purpose of instruction parsing is to extract useful information from spoken instructions and transform it into a form readable by robot,which is crucial for improving the ability of home service robot to identify users’ intentions.However,the existing joint extraction method with shared encodings to parse instructions is not sufficient to model the interaction between tasks,and only action entities and related actions mentioned in instructions can be extracted by these methods.This dissertation aims to improve the ability of parsing instructions for home service robot.The process of executing the instructions of users by home service robot is demonstrated in simulation experiments.Firstly,a joint extraction model with partition encoder is proposed to solve the problem of lacking interaction between tasks.The pre-training model is employed to encode the instructions as feature vectors,which is input into the partition encoder as shared features.The partition encoder is composed of entity gate and action gate.Irrelevant features are filtered and relevant features are classified to obtain task-specific feature vectors.Adversarial training is introduced to improve the robustness of the model.Secondly,users may not fully describe requirements in a single instruction.The same entity that mentioned of different instructions in the same conversation is called coreference.In order to parse intentions in dialogue instructions,a sequence-to-sequence algorithm for task-oriented instructions is proposed to transform unstructured instructions into structured graphs readable by robot.This algorithm is divided into two stages: During the first stage,all nodes and edges of non-coreference nodes in the task-oriented dialogue instructions are parsed by model of Sequence to Sequence,in which the intention of users is translated into the graph structure representation after linear representation.In the second stage,the specific pointing information of all the coreference nodes is obtained by relational graph convolutional network,gated recurrent unit and self-attention method.Finally,simulation experiments are employed to demonstrate the process of parsing and executing instructions for the home service robot.Gazebo,a physical simulation environment in ROS,is utilized to build a home construction.Turtle Bot is selected as the demonstrator,which is implemented with the capabilities of voice recognition,voice synthesis,navigation and obstacle avoidance.The process of voice-controlled home service robot is demonstrationed by Turtle Bot.