Proceedings of the Second Workshop in South East Asian Language Processing

Derry Wijaya, Alham Fikri Aji, Clara Vania, Genta Indra Winata, Ayu PurwariantiWorkshops @ COLING 2025

Abstract

The Second Workshop on South East Asian Language Processing (SEALP), held at COLING 2025, represents a critical milestone in addressing the linguistic diversity of South East Asia. With over 1,200 languages spoken by 680 million people, the region faces significant challenges in digital representation. This workshop showcases recent advancements in commonsense reasoning, ethical instruction-tuning, and specialized linguistic tools for languages including Indonesian, Thai, and Tagalog.

Introduction: Bridging the Digital Divide

South East Asia (SEA) is a region characterized by immense linguistic variety, yet many of its languages remain "underrepresented" or "extremely low-resource" in the global NLP landscape. While national languages like Indonesian, Thai, and Vietnamese serve as lingua francas, the digital sphere is often dominated by English. Furthermore, the historical focus on standardized national languages has frequently overlooked hundreds of regional languages and traditional scripts.

The SEALP workshop aims to bridge this "digital divide" by fostering research that moves beyond generic translation tasks. The proceedings emphasize the importance of cultural context, complex morphology, and ethical AI alignment tailored to the specific social norms of the region.

Technical Methodologies and Innovations

The research presented at SEALP 2025 covers a diverse array of technical approaches designed to handle the unique constraints of SEA languages:

  • Context-Aware Transliteration: The bAI-bAI system introduces context-sensitive algorithms to convert modern Filipino text into the traditional Baybayin script, preserving cultural heritage through technology.
  • Multicultural Model Adaptation: NusaBERT builds upon the IndoBERT architecture, extending its capabilities to be more "multilingual and multicultural." This involves specialized fine-tuning or continued pre-training on regional data to capture local nuances.
  • Resource-Efficient Alignment: The Anak Baik project highlights a budget-friendly methodology for curating "ethical vs. unethical" instructions. This approach is vital for aligning Indonesian Large Language Models (LLMs) without the massive compute budgets typically associated with global labs.
  • Specialized Linguistic Engines: IndoMorph provides a dedicated morphology engine for Indonesian. This addresses the limitations of standard subword tokenizers (like BPE) which often struggle with the complex prefix and suffix systems inherent in many SEA languages.
  • Privacy in Low-Resource Settings: Research into speech content de-identification addresses the critical need for privacy when handling transcripts in data-scarce environments.

New Benchmarks and Datasets

A primary contribution of the workshop is the introduction of several high-quality datasets and benchmarks:

  • Thai Winograd Schemas: A new benchmark designed to evaluate commonsense reasoning specifically within the Thai linguistic and cultural context.
  • Anak Baik Dataset: A curated set of Indonesian instructions used for safety training and ethical alignment of LLMs.
  • NusaDialogue: A dataset focused on dialogue summarization and generation for underrepresented and extremely low-resource languages in the region.
  • Automated Grading Case Studies: A dataset from Thailand exploring similarity-based short answer scoring and sampling strategies for educational technology.

Key Insights: Localization and Linguistic Depth

The proceedings reveal several core themes that define the current state of SEA NLP:

  1. Beyond National Standards: There is a growing recognition that "Indonesian" or "Filipino" NLP is not a monolith. Research is increasingly targeting regional dialects and indigenous scripts.
  2. Localized Ethics: The Anak Baik study suggests that AI safety cannot be a "one-size-fits-all" Western import. Ethical behavior must be defined within the cultural and social norms of the specific region.
  3. The Necessity of Morphology: For many SEA languages, standard tokenization methods are insufficient. Dedicated morphological analysis remains a cornerstone for high-performance NLP in the region.
  4. Efficiency as a Priority: Given the resource constraints in many SEA research institutions, there is a strong emphasis on "low-cost" and "resource-efficient" methods for data curation and model training.

Conclusion and Future Outlook

The Second SEALP workshop saw a significant increase in participation, with 20 technical paper submissions covering tasks from transliteration to question answering. Despite this progress, significant hurdles remain. The dominance of English continues to "pollute" local language datasets, and the vast majority of the 1,200+ languages in the region still lack basic digital support. Future work must focus on scaling these efforts and fostering deeper collaborations between academia, government, and industry to transition these research findings into real-world applications.

Summary updated

BibTeX
@inproceedings{wijaya2025proceedings,
  author = {Derry Wijaya and Alham Fikri Aji and Clara Vania and Genta Indra Winata and Ayu Purwarianti},
  title = {{Proceedings of the Second Workshop in South East Asian Language Processing}},
  booktitle = {Workshops @ COLING},
  year = {2025},
}
Figure