 Good morning, good afternoon, good evening wherever you are and welcome to the AI for Good Global Summit all year, always online. I'm Charlotte Kahn from the ITU, the International Telecommunication Union. Like most of the world, the AI for Good Global Summit has gone digital with weekly online programming, allowing us to reach even more people in 2020. AI for Good perspectives offer expert insights, global visions and shared solutions from the AI for Good community. Today we bring you the second episode of Aetri's AI for Making a Better Tomorrow on the subject of AI-based language learning technology. Aetri is the National AI Research Institute of the Republic of Korea. So let's take a look at a short video about Aetri. Artificial Intelligence Research Laboratory is the main R&D branch of Aetri, the nation's finest general intelligence research institute for shaping our future society. We are in charge of research and development in the field of superintelligence and superperformance to build the foundation of a superintelligent information society co-existing with human-based autonomous intelligence to materialize superperformance computing to overcome the technological limits of existing computers. First, for the realization of a superintelligent information society, the Intelligence Information Research Division is focusing on a language intelligence, voice intelligence, visual intelligence, and big data technology, which are the core technology of the artificial intelligence. Our exo-brain technology, which enables humans to Q&A with computers using natural languages, is being used as a legal inquiry service in National Assembly Library and National Research Council of Science and Technology. Our interactive voice recognition technology is being used as a pilot service to teach AI-based English courses in elementary schools around the country. Our automatic translation Genie Talk technology is applied to solve the language barrier by providing multilingual translation services performed at 2018 Pyeongchang Winter Olympics. Finally, our visual intelligence deep-view technology is applied to track down illegal waste disposal in Sejong City and Seoul. We will continue our research beyond single intelligence technologies into an integrated intelligence technology capable of infinite autonomous growth. Intelligent Robotics Research Division has developed an autonomous driving technology by using AI-based perception, decision, and control technologies. In addition, we can envisage the future of a smart life by developing the core technology of AI Robot that improves a long time and highly intense working environment. With personalized services, Autonomous Unmanned Vehicle Research Department is in charge of researches on the intelligent drone technology that autonomously performs a given mission while flying indoors and outdoors. We also develop an intelligent counter-drone technology that detects, tracks, and neutralizes legal drones. IDX Plus Research Department develops human enhancement and assistive technology that augments human capability by understanding physical and mental conditions. We developed a real-time predictive system to prevent elderly injury due to fallen and a wearable muscular support suit to help them strengthen diminished muscle. The ultimate aim is to develop a cognitive computing technology which understands, learns, and predicts user emotion and behavior based on daily life data. As remarked above, through the Super Intelligence Researches, we will do our best to realize a Super Intelligence Information Society that co-exists with autonomous intelligence centered on humans via a core AI technologies and industrial applications such as autonomous vehicles, intelligent robots, intelligent drones, and the human enhancements. In the future, Atree Artificial Intelligence Research Lab will do its best to advance Korea's artificial intelligence technology to the next level through the development of next generation AI core technology and to achieve the fourth industrial revolution through intelligent innovation in the manufacturing and ICT industries, which are the strengths of Korea. Now let me hand over to our expert host Dr. Miran Choi. Dr. Choi is principal researcher and standardization specialist at Atree. Dr. Choi. Thank you Charlotte. Ladies and gentlemen, I'm Miran Choi from Atree Korea. Thank you very much for joining us in our prospective session. The subject of this session is on AI-based language learning technology, which we have been developing for more than five years. We will introduce AI Language Tutor, which can listen, talk, and teach based on educational contents. The AI Tutor achieves its goal with a high performing speech recognition and dialogue processing technologies. This AI-based language learning technology will be applied to the public education from the next year, and it will contribute to solve the social problems, such as expensive private education costs and communication difficulties of multicultural families. Next, let me introduce the speaker of this session, Dr. Jin Xieofang, as a senior researcher at Atree, has been involved in the research projects on machine translation and dialogue processing for language learning. The technologies that she participated in have been transferred to several companies and are being used in practice. Her current research topic is about the dialogue system based on reading comprehension for language learning. Now, let's watch the presentation. This talk is about AI-based language learning technology. Regional disparity and income inequality cause inequality in education. Further, the education inequality is reported to widen since schools shut down due to COVID-19. Equality of opportunity in English education needs to be insured to reduce the gap in all regions and income and to increase educational equity. Also, it is crucial to support non-face-to-face language education during the COVID-19 pandemic. This is why we started this research. Our goal is to develop an artificially intelligent language tutor which can listen, talk, and teach just like a native tutor. Now, let us briefly introduce ourselves. Atree. Atree is Korea's national research institute that develops the ICT technology to contribute to enhancing industrial competitiveness of Korean enterprises. Recently, Atree intends to play the role of a national AI research institute for making a better tomorrow. Developing an AI-based language tutoring system, AI tutor, is also part of the effort. So, what technology is needed to develop an AI tutor? First, for the AI tutor to listen, spontaneous speech recognition technologies are utilized. Moreover, in order to enable free talks, spoken dialogue processing technologies comes in, including task-oriented and open-domain dialogue technologies. Lastly, it is important to combine AI and educational content optimally for educational purposes. The AI tutor should be able to recognize non-native, spontaneous English speech. Speech recognition is the most important function in English-speaking learning service. Errors are inevitable in non-native speech, ranging from articulation, grammar, to expression. Only if the speech recognizer writes them down as they are, then the system can accurately evaluate pronunciation or speaking fluency, and provide educational feedback on grammar and expression errors in subsequent stages. The AI tutor also needs to assess and give feedback on the pronunciation and fluency. To achieve this, it has to learn native rubric to assess students' speech, compare the native pronunciation with the students for feedback. Now, let's see how AI technology enables the AI tutor to listen. First, for speech recognition, it is important to improve the acoustic model by sufficiently collecting and transcribing the English utterances of non-native Koreans. It is effective to perform these tasks by nationality or by mother tongue. In particular, provided that the training data isn't sufficient, the deep learning acoustic model features performance improvement in proportion to the amount of training data. Pronunciation evaluation is usually conducted by an evaluation model trained by deep learning or regression learning to maximize the statistical correlation between the fluency score manually evaluated by the human expert and the fluency score calculated by the computer. Let's move on to the next skill, talking. Our AI tutor talks freely with students on topics. So, our AI tutor supports off-topic conversations as well as on-topic conversations. On-topic conversations are supported by a task-oriented dialogue system. The system converses with a student on the given topic using its task-oriented dialogue policy. Students can review what they learn by practicing conversation with a system. The system gives the student feedback that includes grammar checking, error correction, and suggestion on proper expressions. Off-topic conversations are supported by an open-domain dialogue system. Off-topic conversations are provided to motivate and intrigue students. Even if the student goes beyond the given topics, the system provides immediate answers to the student's off-topic chats and hence lets the conversation continue. Now we'll check briefly on how the AI technologies enable free talking. The first part is on-topic conversation modeling, which is based on a task-oriented dialogue system. Generally, a task-oriented dialogue system accomplishes one or two goals, such as scheduling and ticket booking. Thus, the knowledge required for a task-oriented dialogue system is relatively limited because it is aligned with one or two tasks. However, a dialogue system for education should help students learn dialogues of various topics. Therefore, it is necessary to construct various dialogue scripts continually. Considering that building knowledge for a task-oriented dialogue system is a high-cost task that requires high expertise, we propose a simple hierarchy format named DialogMap so that the script data can be built easily and quickly and reused efficiently. The DialogMap written for a topic would be automatically converted into the knowledge understandable to the dialogue system, such as knowledge for user intention understanding, dialogue policy rules for dialogue management, and knowledge for dialogue generation. The knowledge generated is used to train each module of the dialogue system, allowing the system to understand the user's intention, manage the conversation, and generate system responses. The second part is off-topic conversation modeling. Off-topic conversation for education purpose differs from the conversation with a general-purpose chatbot. First, the off-topic conversation of an AI tutor should be controllable. To achieve this, we developed Hybrid Open Domain Dialog System, with example-based approach to ensure the controllability and the deep learning-based chatbot to improve the conversation coverage. Another difference is that the persona of AI tutor should be consistent on the same topic while changing in different topics. For instance, an AI tutor can be a box office staff in one dialogue and a school teacher in another dialogue. To maintain consistency of each persona, we classify the persona-related dialogue and proceed the persona-related conversation corresponding to each topic. We have strived to apply AI technologies developed for the last few years to the English Education Service. Four years ago, which was 2016, we launched an AI tutor-based English Language Learning Service in Ulleungdo Island as a pilot project. Ulleungdo is an island at the eastern end of Korea with extremely limited education infrastructure. This experience became a very valuable experience for further services. In 2018, we developed an English Education Service called In-Class with Donga, which is the most influential educational publisher in Korea with over 70 years of history. In-Class is an AI language tutor based on the dialogue selected from textbooks. It consists of three stages, word practice, sentence practice, and debate. Now we're starting a new attempt towards AI Peng Talk Service. Korean educational broadcasting system is developing AI Peng Talk Service with the support of Korean government. Under a pilot project run at 58 elementary schools until last spring semester, we applied to the service our AI tutor technology and collected comments on the AI tutor service from students, teachers, and parents. Let's take a look at the AI Peng Talk Pilot Service. It's a classroom with AI Peng Talk. First is word practice. Is this fun like this? This is a small elementary school in Ul-Tin, Gyeongsang-Puk-Tteok. There are 40 students in this simple rural school. This is the fourth grade classroom. Seven students are studying English with their teacher, Jung. The students review the contents of textbooks, practice correct pronunciation, and compare them with other friends. The students said that studying English became fun with AI Peng Talk Tutor. School class is over. Ji Young, who does not go to the after school, starts studying English at home. This is Yoo Sung's house. Yoo Sung also studies English at home, after school. Since one-on-one customized learning is possible, he reviews what he learned at school. The service is currently being piloted at 58 schools. In 2021, it will be provided to 6,000 elementary schools nationwide. We'll give you a demo of AI Peng Talk Pilot. There are more features under implementation. Today's demo only includes a few of them, such as word, sentence, and conversation level practice, and let's talk dialogue practice. Here are our expectations toward the service. First, it will contribute to providing equal education opportunities to bridge the gap between all regions and incomes. Second, it will expand the students' opportunities to be exposed to and use English within and without the school. This will help every student develop basic communication skills from primary education. Third, it will support non-face-to-face language education during and post the COVID-19 pandemic. We would like to keep our efforts to develop an AI-based language tutor to achieve above goals. Let's start the demo now. Word level practice. Listen and repeat the words. Cook. Cook. AI tutor evaluates students' pronunciation and gives feedback on each word like excellent, good, or try harder. Writer. Writer. Writer. Doctor. Doctor. Teacher. Singer. Dancer. AI tutor also gives final score for students' whole word level practice. Sentence level practice. Listen and repeat the sentence. What do you do? What do you do? I'm a writer. I'm a teacher. I'm a teacher. I'm a singer. I'm a dancer. I'm a dancer. The student is required to repeat the sentence. I'm a dancer. AI tutor also gives final score for the sentence level practice. Conversation level practice. Listen and repeat the conversation. Listen to the whole dialogue. Are you a singer? No, I'm not. What do you do? I'm a dancer. Are you a singer? You are a singer. AI tutor gives feedback on each sentence. No, I'm not. What do you do? What do you do? I'm a dancer. I'm a dancer. AI tutor gives final score for students' whole conversation level practice. The student is starting to practice another conversation on the same topic. Teacher? No, I'm not. What do you do? I'm a writer. Are you a teacher? You're a teacher. No, I'm not. No, I'm not. What do you do? What do you do? I'm a writer. I'm a writer. Let's Talk enables episode-based free talks covering both on-topic and off-topic ones. The first episode is about Pengsoo, the AI tutor. The student must ask Pengsoo questions about his name, where are you from, and where do you live to complete the mission. Peng-ha! All about myself. Ask me questions. What's your name? Great. Pengsoo, let's be friends. What's your dream? What's your dream? I want to be a superstar like Beats. Where are you from? Great! Second mission, where are you from, success. I'm from the South Pole. I want to be a superstar. Where do you live? Great! Final mission, where do you live, success. I live the easiest now. I'm from the South Pole. All missions complete. Ask me five questions or say goodbye. Do you like them? AI Pengsoo can handle them. No, I don't. I like dancing. Ask me for more questions. Are you happy? Couldn't be happy. Ask me three more questions. Can you swim? Of course, I love to swim. Ask me to more questions. Sorry, I don't get it. Ask me again. Okay, what is your last question? Goodbye. It's time to go now. See you next time. Peng-fa! Another episode in Let's Talk. The mission is buying fruits. Students practice English with AI Pengsoo through role-plays. May I help you? First mission, say the name of the fruit success. How many? How many? The student tests the calculation ability of AI Pengsoo by ordering 100 apples. Okay. Here we go. I have no money. Don't give up, just check the hint. How much is it? Great. It's 100,000 won. Thank you, please come again. Bye. Free Talks encourages students' participation and makes studying English enjoyable. Thank you for your presentation, Dr. Wang. I hope you have a good achievement in applying AI technology to language learning for the online education. Now, I will ask three questions on your presentation for the audience. My first question is you just introduced AI Language Tutor. Can I test it by online connection or by downloading the AI Language Tutor application? The quick answer is yes. You can download EBS AI Pengto from Korea Android App Store. It's still a trial version with only part of functions present. But we can expect direct version from next year. As we introduced in our presentation, we will introduce AI Tutor and external education companies to provide final service. That's the case with AI Pengto. The second question is if someone in other countries wants to use this technology, what should the person do? Well, if you're just an end user like a teacher or student who wants to teach English, you can use the app provided by the current app store directly. If it is the other way around, your company that wants to provide services in your country can do a lot more. For example, in your country there may be unique English pronunciation or special English teaching content. Therefore, it is better to optimize the system for your country. You can use your pronunciation data to refine your speech recognition model or developing your own educational content for the dialogue system. In the purpose of use it for other language education such as French, German, Chinese or Japanese, you'll need to have your own training data and educational content for the language to retrain our system. Your answer will help a lot for people who want to test your system. Now my last question is can you tell me about future plans on this project? We have transferred our English and also Korean language learning technology to several companies including EBS and Dona Manson in today's presentation. With these tech transfers we hope that students will have easier access to various online language learning content both at school and at home. The research on technology also continues. In terms of speech recognition we are starting how a system can travel high performance by learning only a small amount of speech data. In the case of dialogue system we are starting an end-to-end technology that can automatically extract and learn conversation knowledge from traditional language education content without extra effort for dialogue map. Through these studies we hope that AI language learning technology will be applied more quickly and easily to more languages Thank you Dr. Huang for your detailed explanation. Okay I hope you all have enjoyed the episode 2 of AI for Good Perspective presented by Atri today. Thank you for your participation and let me hand back over to Charlotte for the closing. Thank you Dr. Choi and Dr. Wang for the presentations and inspiring discussions. Please join us again on the 3rd of November for another fascinating insight into Atri's advanced work on AI in the Republic of Korea when the subject will be AI for security and privacy.