Personalized Conversational Agent for Second Language Learning
For thousands of years humans have developed features and technologies that would facilitate the life for many. That meant a compromise that these facilities are not perfect for each person but would work for most. Each individual is unique in the environment and resources available to them. But some people differ more than the others from the norm of the society, who shape the minorities. Those solutions, although compromised, would not work on these other people and they would be left out. However, this is in violation of modern human society rights. All humans should have equal rights and language rights are part of them.
Project Description
Language is a fundamental defining feature of human beings and so, is the right of understanding, being understood, availability of content, learning a new language. Emergence of global village has brought up the need to learn new languages for many people. Digital technologies provide an excellent opportunity for language learning. However, these technologies, like other human solutions, are mostly set of compromises for the majority, which is not a perfect solution for everyone. Therefore, causes exclusion, and reduces opportunities of certain society members. One step towards solving this issue is to consider all individual’s differences with Adaptive Intelligent Systems and personalize the learning experience. A general system or application that is same for all, is not optimal due to different needs of each individual, and resource limitations would not allow having a teacher for each person.
Conversation is one of the most important dimensions of the language. We have used the GPT-3 language model to develop a smart conversational agent for grammar learning and conversation practice in English. The conversation starts with a prompt from the agent asking the user a simple question to start the discussion. Each user’s answers go along a pipeline: First, the response is passed through two grammatical error detection systems to spot if there are any mistakes in the sentence, and if yes, where and what is the correct way of saying it. Second, if the sentence is correct, agent will produce a relevant response, otherwise it will enter a correction loop. One of the different correction feedback types (No feedback, Corrective, Informative, Corrective+Informative, Elaborative) then is given to the user. The user rates each response of the agent in <good> or <bad>. These correction feedbacks initially are randomly provided. Currently, we collect crowdsourced data to train a baseline model which can decide when and how to provide which type of correction feedback. In the next stage, using the methods of continued learning in machine learning we will adapt the baseline model for each individual using the system based on their personal experience. A language test is administered in the end of the conversation from user’s own mistakes to optimize the model’s responses during the next iteration.
Such personalized learning experience has considerable advantages: 1) Meeting everyone’s needs individually, they assure inclusion. Adaptive agents are not same for everyone, but unique for each person. Therefore, they would work for minorities like people with language deficits, people with social anxiety, ASD, and ADHD; 2) Optimizes user’s learning by emphasizing on automatically detected weak points, thus saving time, and increasing learning efficiency.
Researchers
