Identifying negative language transfer in learner writing: using syntactic information to model structural differences

Carrie Demmans Epp, Academic Team - UA

18 August 2021


Language transfer is a phenomenon in second language acquisition in which learners reuse rules from their first languages when communicating in a second language. When the first language rules differ from the second language ones, reusing them results in errors. Those errors are called negative language transfer. Although the negative language transfer phenomenon is well-researched, there have not been many projects that use computational language modelling to detect this phenomenon.

In this project, we applied language modelling techniques to the syntactic information of learners’ first and second languages. The language models created were then used to identify first language structures that were present in errors made by second language learners. We developed this method to understand whether representing language structures using syntactic information could be useful to detecting negative language transfer errors.

Once negative language transfer identification models were created and their capabilities were established, the project started to focus on presenting negative language transfer feedback to language learners. The purpose of this step in the project is to make learners aware of the negative language transfer phenomenon. We hope that by explaining the phenomenon to learners, they will be able to better differentiate between first and second language rules.

The development of language models using syntactic information to detect negative language transfer was Leticia’s master’s thesis. Nicole was fundamental in the evaluation of those models as she annotated language-learner errors with negative language transfer information. Carrie supervised the project and guided the team. Jiahua will be helping to evaluate how learners respond to the developed technology.

The language structure representations that we used to detect negative language transfer achieved reasonable results. For some types of errors, these representations were successful in detecting when a learner error was related to a transferred first language rule. These results indicate that syntactic information, such as part-of-speech tags, is useful for detecting negative language transfer between languages that are structurally distant.

Team members: Carrie Demmans Epp, Leticia Farias Wanderley, Nicole Zhao, Jiahua Liu