Open AccessExtract Executable Action Sequences from Natural Language Instructions Based on DQN for Medical Service Robots
Author(s) -
Fengda Zhao,
Zhikai Yang,
Xianshan Li,
Dingding Guo,
Haitao Li
Publication year - 2021
Publication title -
international journal of computers, communications and control
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.422
H-Index - 33
eISSN - 1841-9844
pISSN - 1841-9836
DOI - 10.15837/ijccc.2021.2.4115
Subject(s) - executable , computer science , robot , action (physics) , service (business) , artificial intelligence , natural language , core (optical fiber) , natural (archaeology) , simple (philosophy) , natural language processing , human–computer interaction , programming language , telecommunications , physics , economy , archaeology , quantum mechanics , economics , history , philosophy , epistemology
The emergence and popularization of medical robots bring great convenience to doctors in treating patients. The core of medical robots is the interaction and cooperation between doctors and robots, so it is crucial to design a simple and stable human-robots interaction system for medical robots. Language is the most convenient way for people to communicate with each other, so in this paper, a DQN agent based on long-short term memory (LSTM) and attention mechanism is proposed to enable the robots to extract executable action sequences from doctors’ natural language instructions. For this, our agent should be able to complete two related tasks: 1) extracting action names from instructions. 2) extracting action arguments according to the extracted action names. We evaluate our agent on three datasets composed of texts with an average length of 49.95, 209.34, 417.17 words respectively. The results show that our agent can perform better than similar agents. And our agent has a better ability to handle long texts than previous works.