During the pandemic, Cognitive Science’s Vera Tobin received a message from one of her undergraduate students seeking a research opportunity. The student, Alekhya Yadavalli, and her brother Aditya wanted to combine their cognitive science and computer science skills with a linguistics expert to refine their ideas and collaborate on a project. What started out as an exploratory idea built up to be much more.
“For a part of the time, Alekhya was studying abroad in Oxford while Aditya was in Hyderabad and I was in Cleveland, so making the time zones work out was a bit of an exciting adventure in itself,” Tobin said.
Constructing the collaboration
The trio wanted to combine Alekhya’s knowledge of cross-linguistic corpora, Aditya’s software skills and Tobin’s ability to analyze the results in light of their linguistic significance. They built machine learning models of second language acquisition to investigate what a machine will carry over from the first language to the next.
“Many existing projects are centered around English, the language with the most data. Instead, we wanted to test models using other languages that are less tested and ask, ‘how can we do more with less?’”
Using child-directed speech (CDS) allowed them to work with language tailored to beginners, with more repetition, simpler vocabulary and other structural features that support language learning. Because CDS shares these distinctive features across languages, it was used to give the machines a general boost in language learning across the board by setting up their models to learn an initial, smaller language and then learn English as their second language.
Results of the research
Just like humans, machines find it harder to acquire a second language that’s more distant from its first language.
“The machines made a kind of human-like mistake that is the result of ‘negative transfer,’ which is what happens when something you learn in one context actually interferes with your acquisition of a new skill or concept,” says Tobin.
Tobin’s research shows that it’s meaningful to not only measure transfer, but to be able to identify negative transfer and develop models that go further than concluding that “poor performance = less transfer.”
Additional results show that some qualities of CDS are crucial to aspects of human language learning that text-based models aren’t able to incorporate, such as detecting word boundaries in speech and identifying what words refer to in the real world.
Looking ahead
The team may start making their current models more robust to have a more fleshed out approach for finding specific error patterns between tests. Additionally, they aim to develop systems that are able to test for and distinguish the difference between less positive transfer and the presence of negative transfer.
What began as students approaching a professor with an unfinished research idea resulted in the students presenting their published study at the highly regarded Association for Computational Linguistics conference in 2023. It was also the first time that the three had met in person after conducting all of their research remotely.
“It was a wonderful and strange way to meet, but very easy in the warm connections we had already established within our team,” Tobin said. “Ten years ago, no one would imagine research being done like this.”