Learning Low-Resource Languages Through NLP-Driven Flashcards: A Case Study of Hokkien in Language Learning Applications

Tai Zhang, Lucie Yang, Erin Chen, Karen Riani, Jessica Zipf, Mariana Shimabukuro, En-Shiun Annie LeeNAACL 2025

TL;DR

This paper introduces LangLearn, an open-source framework that uses AI-driven flashcards to help users learn low-resource languages (LRLs) like Hokkien. By combining language-agnostic AI models with personalized content, LangLearn aims to make LRL acquisition more accessible and engaging, addressing the current gap in digital language learning tools for underrepresented languages.

What Problem Does This Paper Address?

The proliferation of language learning apps like Duolingo has made language acquisition more accessible, but these platforms predominantly focus on popular languages. This leaves low-resource languages (LRLs), such as Taiwanese Hokkien, significantly underrepresented. LRLs often suffer from a lack of training data for AI models, limited written materials, and dialectal variations, making it difficult to develop effective digital learning tools. This creates a critical gap in technological support for speakers of these languages, many of whom are facing a decline in usage. While personalization in language learning has been shown to boost motivation, it hasn't always improved vocabulary acquisition itself.

Methods

LangLearn is built as a modular, AI-driven framework designed for autonomous learning. Its core components include:

  • AI-Enhanced Flashcards: The system generates flashcards with six key elements: the word definition, corresponding Hanji characters, romanization with numeric tones, audio pronunciation, a sample sentence, and an AI-generated contextual image.
  • Language-Agnostic Approach: The framework is designed to be adaptable to various LRLs by integrating language-specific AI models.
  • Modular Architecture: This allows for easy expansion and integration of new features, such as Automatic Speech Recognition (ASR) for pronunciation practice.
  • Hokkien-Specific Models: For Taiwanese Hokkien, the system leverages pre-trained models for translation, transliteration, and text-to-speech, all integrated into a flashcard generation pipeline.
  • Text-to-Image Generation: An AI model generates contextual images to aid understanding and retention.
  • User-Centered Design: Features like customizable decks, quizzes, and a community rating system for AI-generated content are included to enhance user engagement and improve accuracy.

Datasets

The paper does not explicitly mention specific datasets used for training the underlying AI models. However, it relies on pre-trained models for Taiwanese Hokkien, including translation, transliteration, and text-to-speech models, as well as the OpenAI API for text-to-image generation. The study's evaluation used data collected from a preliminary user study involving 39 participants.

Key Insights and Ideas

  • AI as a Bridge for LRLs: LangLearn demonstrates how AI can be leveraged to bridge the gap in digital resources for LRLs, even with limited data.
  • Motivation through Personalization: While direct vocabulary learning gains from personalization weren't the primary focus, the study suggests that personalized content (like contextualized definitions and images) significantly improves learner motivation.
  • Multi-Modal Learning: The integration of text, audio, and AI-generated images in flashcards provides a rich, multi-modal learning experience that can aid comprehension.
  • Community-Driven Improvement: Incorporating user feedback and a rating system for AI-generated content is crucial for addressing the inherent inaccuracies of current genAI models and improving the system over time.
  • Extensibility is Key: The modular design of LangLearn is vital for its long-term viability, allowing it to be adapted for other LRLs and incorporate new features.

Key Results

A preliminary user study with 39 participants learning Hokkien yielded promising results:

  • Usability: 82% of users agreed the app was easy to use.
  • Engagement: Many users enjoyed the novelty of contextualized definitions and AI-generated images.
  • Intuitive Use: The majority of responses indicated the app was intuitive and easy to interact with.
  • Learning Effectiveness: While initial feedback on the platform being "helpful for learning Hokkien" was mixed (with some users noting the lack of audio in early versions), the updated version with speech and pronunciation features is undergoing further study.

Limitations and Future Work

Limitations:

  • AI Accuracy: The paper acknowledges that AI-generated content (especially images and contextual sentences) can have accuracy and relevance issues, which are being addressed through user feedback and community ratings.
  • Study Scope: The preliminary study was conducted over three weeks with undergraduate students, and the initial version of the app lacked audio components, which were later added.
  • Longitudinal Data: The study noted a decrease in survey responses over the weeks, a common challenge in longitudinal studies.

Future Work:

  • Expand Language Support: Integrate more translation models to support a wider range of LRLs.
  • Enhance Pronunciation Practice: Implement ASR-based pronunciation practice to further emphasize the spoken nature of many LRLs.
  • Grammar and Verb Conjugation: Develop flashcards that go beyond vocabulary to include grammar and verb conjugation practice.
  • Comparative Study: Conduct a future study to compare the effectiveness of simple flashcards (word + pronunciation) versus enhanced flashcards with additional genAI components (sample sentences and images).
  • Notification System: Explore generating notifications and contextualized learning suggestions for on-the-go learning.

Summary updated

BibTeX
@inproceedings{zhang2025learning,
  author = {Tai Zhang and Lucie Yang and Erin Chen and Karen Riani and Jessica Zipf and Mariana Shimabukuro and En-Shiun Annie Lee},
  title = {{Learning Low-Resource Languages Through NLP-Driven Flashcards: A Case Study of Hokkien in Language Learning Applications}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure