Konnichiwa, Chatbot! Students to Leverage AI for Fast-Track Language Learning

woman at computer

Custom tools simulate on-demand conversations with native Japanese speakers

by MaryAlice Bitts-Jackson

Imagine world-language students practicing conversations with skilled speakers whenever and wherever they wish. And imagine that they could submit automatic transcripts to their professor—in other words, turn in their homework—with just a click of the mouse. These futuristic conceptions are emerging realities at Dickinson. An exciting academic-tech project aims to enhance teaching and learning of the Japanese language through the power of artificial intelligence (AI).

The project is led by Todd Byrant, language technology specialist, and Akiko Meguro, senior lecturer in Japanese. It harnesses AI chatbots—​​computer programs that mimic human conversation—to simulate on-demand conversations with a native Japanese speaker. Four chatbots are developed and in testing, with more on the way.

Bryant was inspired by research at Northeastern University’s World Language Center indicating that nearly all students using world-language chatbots reported positive experiences as well as enhanced speaking ability and comprehension. The challenge? Customization is required—a heavy lift. Bryant and Meguro were eager to see how AI might add to teaching and learning Japanese, however, so they got to work.

‘Writing a story’

Meguro developed the scenarios. Bryant worked with student-programmer Leah Goldberg ’23 to lay a foundation for beginner-level-Japanese chatbots, to be accessed on a dedicated webpage through Dickinson’s website. Noah Lape ’26 joined the team as student-programmer last spring.

The students developed the chatbots using Rasa, an open-source platform used to create AI chatbots. Each bot is programmed to enact a specific conversation--a visit to the doctor, say.

“It’s called ‘writing a story’ in Rasa,” says Bryant. “You have a structured conversation and the directions you expect it to go.” Custom code, written by the student-researchers, handles user input that that Rasa may not expect and creates additional functionality, such as creating a transcript or sending a notification to the professor when the conversation is complete.

Hold the anchovies!

Like chatbots you might use to order pizza through an app, these bots work well with a limited number of options, like the size of the pie, the kind of crust and the toppings—or for simple responses, such as that the weather is cloudy, sunny, snowy, rainy, cold or hot. This fits the needs of beginning language learners: In fact, the more structured the conversation, the easier it is to keep them on track. And if a user delivers an incorrect response, the bot can easily direct them to a relevant page in their text.

But as language skills advance, conversational possibilities multiply, and limited options are—well, limiting. While that pizza bot is great at putting in an order, you can’t ask it about the health benefits of whole-grain crust and expect a helpful reply.

Enter a new, heavier-hitting tool that’s been taking the world by storm.

Leveraging ChatGPT

Last November, OpenAI launched ChatGPT, a large-language chatbot that enables more open-ended, rather than “if/then” scripted conversations. This opened a door to chatbots for more advanced language learners as well.

For these learners, the Dickinson team uses Rasa to provide a basic bot structure and ChatGPT API to generate more robust “thinking” and responses. The key  is to define the bot “speaker’s” role as well as the conversational tone.

One exercise prompts students to imagine meeting a Japanese roommate for the first time. For example, “we basically say [within the prompt sent by the API]: ‘Hi, ChatGPT. You’re a super-friendly roommate from Nagoya,’” Bryant explains. The API delivers the student’s typed text to ChatGPT with those parameters, and ChatGPT’s responses are delivered back to the web interface. Additional code addresses what to do if the conversation goes off track.

It takes a lot of time to code—and quite a bit of fine-tuning by Meguro. If, say, most of what ChatGPT “knows” about Japanese comes from being fed text from Japanese newspapers, then its output will sound a lot like news copy—formal and stiff. As she tests each bot, Meguro keeps an eye out for overly formal, awkward language, phrasing and tone, and asks the team for adjustments.

Ahead of the pack

Bryant notes that while private-market concerns are developing world-language chatbots, this sort of project is rare in higher education, placing Dickinson ahead of the curve. And with OpenAI’s Sept. 25 announcement that users will soon be able to hold voice conversations through ChatGPT, there’s a good chance that, in time, students will be able to practice speaking in Japanese, as well as writing, through the Dickinson chatbots.

Meguro plans to introduce the chatbots to students in her Japanese 201 class and collect feedback this semester. Plans are to roll out the chatbot next year.

While most Japanese-language students must wait a bit longer to access these tools, two Dickinsonians have already benefitted from the project: student-programmers Goldberg and Lape.

“It’s been a thrilling introduction to machine learning,” says Lape, who’s also sharpened skills in both Japanese and computer science through this venture, coded with Rasa, used and prompt-engineered ChatGPT API, and developed a web interface. “I'm still learning new techniques and improving my problem-solving abilities every day.”



Published October 9, 2023