Abstract—Commercial chatbots such as Apple’s Siri, Microsoft’s XiaoIce, Amazon’s Alexa, Jingdong’s JIMI, and Alibaba’s Alime, have some great prospective in applications such as hosting programs, writing poetry, providing pre-sale consulting and after-sales service in E-commerce, and providing virtual shopping guidance. However, in most cases, existed chatbots in the world are neither designed specifically for children, nor suitable for children, especially for children with ASD (autism spectrum disorder). In order to develop chatbots that are suitable for children with ASD, the present study firstly adopted an open source chatting corpus containing more than 1.7 million question-and-answer Chinese sentences of chatting histories involving children in many cases, and screened out more than 400,000 ideal chatting sentences for model training. Then a generative-based method combing Bi-LSTM and attention mechanism with word embedding based on deep neural network was adopted to build a general Chinese chatbot. The quality evaluation results indicated that our chatbot can successfully intrigue participants’ interest and made them understand it well. The chatbot also showed its’ great potential for using in the conversation-mediated intervention for Chinese children with ASD. Index Terms—Chatbots, Autism Spectrum Disorder, Natural Language Processing, Generative-based Method, Deep Learning I. INTRODUCTION Autism is a developmental disorder and 1 in 160 children has an ASD worldwide [1]. Common symptoms of children with ASD are deficits in social and language ability, as well as repetitive behaviors and narrow interests [2]. Fortunately, research indicates these abilities can be significant improved by early intervention [3]. In recent years, researchers show increasing interests in using computer assisted technology (CAT) and rehabilitation robots as interventions for children with ASD. Computer assisted technology has long been recognized as an effective treatment for children with ASD. [4] A dozens of research [5][6][7] report that autistic children normally feel comfortable in an environment which is predictable and avoid directly contact with human. Therefore, children with ASD take delight in interaction with computer. Computer assisted therapy for children with ASD contains a wide range of applications. Previous interventions mainly focus on using Manuscript received August 15, 2019. This work was supported by the Shenzhen Science and Technology Innovation Commission under Grant JCYJ20170410172100520. Xuan LI is with the College of Software, Xi’an Jiao Tong University, China (e-mail: lixuan064@ stu.xjtu.edu.cn). Huixin ZHONG is with the Department of Computer Science, the University of Bath, United Kingdom. (e-mail: huixinzhong@outlook.com). computer or mobile devices to conduct emotion recognition training [8][9], language learning [10] and social behavior education [11]. Robot assisted treatment for children with ASD also achieved tremendous success and showed its unique value in clinical practice. Existed rehabilitation robot- based interventions generally concentrated on improving social skills of autistic children through joint attention [12] [13], learning imitation and turn taking [13]. However, none of these interventions apply chatbots which can simulate real communication in real society to enhance the social and communication skills of children with ASD. Since children with ASD have linguistic difficulties [14] and commonly reluctant to communicate with others [15], early intervention mainly focuses on language teaching. However, researchers and therapists should bear in mind that the final goal of intervention for children with ASD should be facilitating autistic children to return and to apply skills which they learned from interventions to a real world. Therefore, to involve chatbots which could simulate human to human communication would be a valuable and important attempt to reach this goal. It is necessary to apply chatbots in CAT or robot-based intervention. However, existed chatbots lack adaptability to language comprehension of autistic children. Therefore, the present study designs a general Chinese chatbot based on deep learning and some adjustments were made according to autistic children’s language habits and understanding abilities. For example, autistic children with language barrier generally have pragmatic impairments [16] [17] [18] such as lacking of response or having difficulty in sustaining a conversational topic. Experts consider that the pragmatic impairments may relate to the difficulty in language comprehension [19]. Therefore, in the design of chatbot, we keep responsive sentences as short and simple as possible to make it easier for autistic children to understand. Besides, we conduct pre- training for the word embedding and therefore the imagination of our chatbot has been significant improved and this allow the chatbot generating appropriate response even to single input words. Considering the ages of children, our chatbot selects relatively pure corpus, which do not contain sexual or violent languages and it also makes our chatbot to have an optimistic personality. To our best knowledge, our chatbot is the only general Chinese chatbot designed for meeting the special needs of children with ASD. The rest of this paper are arranged as follow. Firstly, we Bin ZHANG is with the School of Software & Microelectronics, Peking University, China (e-mail: zhangbinxy@pku.edu.cn) Jiaming ZHANG is with the Institute of Robotics and Intelligent Manufacturing, The Chinese University of Hong Kong (Shenzhen), China (e-mail: zhangjiaming@cuhk.edu.cn). He is also with Shenzhen Institute of Artificial Intelligence and Robotics for Society, China presents for the joint first author, *presents for corresponding author. A General Chinese Chatbot based on Deep Learning and Its’ Application for Children with ASD Xuan LI, Huixin ZHONG, Bin ZHANG 2 , Jiaming ZHANG*