Towards Language Learning From Passive Exposures To In-Context Examples

This thesis explores innovative methods for language learning through passive exposure to in-context examples. We aim to reduce the discipline, willpower, and time investment required to learn a new language by integrating language learning into existing daily habits. The fundamental idea is to translate some parts of the text that the user is reading. In a first work we present the software components necessary to implement this idea, discuss efficient learning strategies for smart word selection and evaluate the learning paradigm in a comprehensive user study. To improve the learning experience and intuitively guide the reader towards a correct understanding of the foreign language demonstrations, we design a novel system that can translate text into a series of semantic images. In a cloze-test-based user study, we find that our visual semantic cues significantly increase the chance of correctly guessing a masked word. We hope that in practice, our work will optimize the process of learning from passive exposures by reducing ambiguity. Finally, we discuss a shortcoming of translating words in place: We are limited to teaching vocabulary to our users that occurs naturally in the text they are reading. To overcome this problem, we design an NLP pipeline that uses generative AI to rewrite and extend the text read by our users.