Students on computers

Meet the next AI-powered “Language Buddy”

When it comes to language learning, communication is the ultimate goal. Improved linguistics, comprehension, grammar, vocabulary and pronunciation are just some of the benefits of learning a language through speaking. 

“The more opportunities you have to communicate in a foreign language, the more you advance and have fun with it,” explained Christiane Reves, an assistant teaching professor of German in Arizona State University’s (ASU) School of International Letters and Cultures. 

For communication to take place, you need a partner. 

However, regular interaction isn’t always possible for students in ASU Online language courses; diverse student body learning needs and scheduling demands can make it challenging to hold synchronous instruction and peer virtual meetups. “Yes, students can use different apps, but opportunities for tailored, real-life conversation can be very limited,” said Reves.

Reves and colleagues in her department think that “Language Buddy” – a GPT that they created in ChatGPT Enterprise as part of the university’s AI Innovation Challenge – could be the solution to getting students to practice their language skills. Powered by generative artificial intelligence (AI), Language Buddy will allow students to participate in conversations at their language level — anytime, anywhere.
“If AI is flexible and acts in human ways, what can we ask it to do so students can practice their language skills,” asked Reves.

Since before the pandemic, Reves and her colleagues have long been early adopters of technology to improve current teaching and learning strategies. Currently, students are meeting with the instructors and peers via Zoom, and are recording themselves speaking on the Flip app and commenting to each other asynchronously. “We see [technology] tools as a positive enhancement of what we can do,” said Reves.

Experimenting with prompt engineering

New to generative AI, Reves explained that when they first gained access to ChatGPT Enterprise, the team was exploring its general features and what it can and cannot do. For the past month, Reves and her co-collaborators have been training the custom GPT, within ChatGPT Enterprise, using the dictation function to create a space for students to practice free speaking and interacting directly with the bot. They are testing a potential GPT for German students – Germany has a large amount of content available online, making it easier to train the bot – and, if successful, the goal would be to create GPTs for additional languages.

Currently, they are exploring how well the bot can interact with students based on the prompts given, and training the technology to be more human-like and to interact at an appropriate language level.

Ideally, the team would like the GPT to be able to have conversations in German with beginner-level German students (based on the course content), generate a transcript of the conversation and provide tailored feedback for the student on the conversation.

“We’re trying to encode expertise into the bot so it acts on the respective language level,” said David Parks, Learning Support Services Head, School of International Letters and Cultures. “ChatGPT has no idea what you’re trying to do, so we’re really trying to reign it in, in the context of the classroom.”

While Reves and Parks take on experimenting in the mobile app and practicing speaking with ChatGPT Enterprise, Kristin Elwood, Director of Digital Initiatives and Instructional Design Clinical Assistant Professor, School of International Letters and Cultures, is focused on finetuning the prompts. Elwood shared how important tone is when training a bot. “You need to be positive and give some feedback,” said Elwood.

Initial findings

So far, the team has found that the bot can hold conversations, but after a few minutes of interaction, the bot’s vocabulary becomes too complex for the average novice language learner to continue with the conversation. It is critical for the team to establish and define language levels within ChatGPT Enterprise, as the course work follows the American Council on the Teaching of Foreign Languages (ACTFL) Proficiency Guidelines. “We are establishing language levels with the bot, figuring out how to train the bot to speak – and evaluate – at novice and intermediate levels,” said Reves.

If the team figures out how to best train the bot to conduct level-conscious conversations that are timely to what students are learning in the classroom, it not only allows unlimited practice but would also be a major benefit to students who are shy. “With AI, people will feel less judged,” said Elwood. “This gives them advantages of wanting them to speak, but if it gets too complicated too quickly, the students will shut down.”

Next steps for Language Buddy

The team received an extension for the summer 2024 round of the AI Innovation Challenge to continue their work. “We are excited that students will already have the opportunity to try the Language Buddy this summer,” said Reves. “Their feedback will help us with the next stage of development.”

Participating in this cohort will mean that the team can test their bot with students for the first time in Reves’ German 101 class this summer. Reves has “no hesitation” to share the AI-experience with students, but students need to keep in mind that “it’s ready for use, but not ready for assessment,” said Reves.