Using Speech and NLP Resources to build an iCALL platform for a minority language, the story of An Scéalaí, the Irish experience to date

Neasa Ní Chiaráin, Oisín Nolan, Madeleine Comtois, Neimhin Robinson Gunning, Harald Berthelsen, Ailbhe Ni ChasaideACL 2022

TL;DR

This paper introduces An Scéalaí, an intelligent Computer-Assisted Language Learning (iCALL) platform designed to revitalize the Irish language, an endangered minority language. By integrating speech and NLP technologies with a learner corpus, An Scéalaí aims to provide a holistic learning experience and gather valuable data for understanding language acquisition.

What Problem Does This Paper Address?

The paper addresses the critical challenge of language endangerment, specifically for Irish. Traditional language learning methods for Irish are often criticized as outdated, and learners lack access to native speakers and authentic language use. Furthermore, the digital world presents a "digital timebomb" for minority languages due to a severe deficit in speech and language technologies. This gap hinders the language's use in modern contexts and its transmission to future generations. An Scéalaí aims to bridge this gap by leveraging emerging linguistic resources and technologies to create a modern, engaging, and effective learning tool.

Methods

An Scéalaí is built on a modular architecture, integrating various linguistic and speech technologies through a REST API. Key components include:

  • Text-to-Speech (TTS): Utilizes the ABAIR TTS synthesizer, offering HMM- and DNN-based synthesis across three Irish dialects. It also provides timing information for live text highlighting, syncing audio with text.
  • Grammar Checker: Integrates An Gramadóir, a microservice-based grammar checker, with an added algorithm for common spelling errors related to vowel agreement.
  • Voice Recording: Allows students to record their own speech and compare it to the synthesized native speaker model, aiding pronunciation practice.
  • User and Content Management: A robust system for managing users (students, teachers, admins) and content, ensuring data is harvested for the learner corpus.
  • Learner Corpus (An Corpas Cliste): An engagement system logs user interactions (grammar checks, TTS usage) to create a rich dataset for analyzing language acquisition patterns.

The platform's development involved a cycle of design, implementation, testing, and redesign, with continuous collaboration with the Irish language learning community.

Datasets

The paper does not explicitly detail specific datasets used for training the underlying speech and NLP models. However, it highlights the creation and use of An Corpas Cliste (The Intelligent Learner Corpus) as a crucial dataset. This corpus is built from user interactions within the An Scéalaí platform, capturing learners' linguistic output and engagement with the tools. This corpus is intended to guide future platform development and research into language acquisition.

Key Insights and Ideas

  • Holistic Language Learning: The platform is designed to simultaneously develop all four language skills (writing, speaking, listening, reading) and reinforce them, moving beyond traditional text-based approaches.
  • Bridging the Digital Divide: An Scéalaí demonstrates how speech and NLP technologies can be adapted for minority languages, actively combating the "digital timebomb" and ensuring their relevance in the digital age.
  • The Power of Learner Corpora: The creation of An Corpas Cliste is a significant contribution, providing invaluable data for understanding how learners acquire an endangered language and how technology can personalize learning.
  • Community-Driven Development: The success of An Scéalaí is heavily attributed to the active participation and continuous feedback from the Irish language learning community, including teachers and learners.
  • Irish Orthography Challenges: The paper notes that the written form of Irish is "opaque," making the link between sounds and spelling complex. The integrated TTS and voice recording features are crucial for learners to grasp this phonic structure.

Key Results

The paper reports overwhelmingly positive feedback from both learners and teachers.

  • Learner Responses: Over 5 months, 1793 learners registered. A significant majority (over 85-90%) reported that the platform enhanced their learning, improved their confidence, and expressed a desire to continue using it. Terms like "revolutionary" were used in open-ended comments.
  • Teacher Responses: 85 teachers registered, with high percentages finding the platform useful as a management system and for future use. The ease of use was generally rated positively.
  • Learner Corpus Size: By the time of writing, the corpus contained 42,542 stories, totaling over 5.5 million words.

Limitations and Future Work

  • Limited TTS Usage: The analysis revealed that the synthetic speech output was relatively underutilized. This is attributed to the novelty of "proof-listening" and its placement in a separate tab from the main writing interface. The system has been redesigned to integrate TTS more prominently.
  • System Robustness: As user numbers grew, particularly during the pandemic, the system experienced crashes. Ongoing efforts are focused on improving robustness and requiring continuous technical support.
  • Future Directions:
    • Increasing Irish language content, potentially using avatars for feedback delivery.
  • Developing independent grammar checking modules for more nuanced error detection (e.g., genitive case marking).
  • Integrating an interactive chatbot for practicing irregular verbs and a story starter for writing prompts.
  • Redesigning the system for mobile/tablet use, recognizing the importance of mobile access for minority language contexts.
  • Expanding the platform to be language-independent, allowing it to be cloned for other endangered languages by swapping out language-specific resources.

Summary updated

BibTeX
@inproceedings{chiarain2022using,
  author = {Neasa Ní Chiaráin and Oisín Nolan and Madeleine Comtois and Neimhin Robinson Gunning and Harald Berthelsen and Ailbhe Ni Chasaide},
  title = {{Using Speech and NLP Resources to build an iCALL platform for a minority language, the story of An Scéalaí, the Irish experience to date}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure