TL;DR
This paper introduces GuyLingo, a comprehensive corpus of Guyanese Creole designed to boost Natural Language Processing (NLP) research for this low-resource language. By creating this dataset and a translation tool, the authors demonstrate the potential for NLP advancements to accelerate the formal adoption of Creole languages in the Caribbean.
What Problem Does This Paper Address?
While major languages receive significant attention and resources in NLP, a vast number of smaller, indigenous, and regional languages are computationally underserved. The Caribbean, often labeled "English-speaking," is home to numerous Creole languages that thrive alongside English but lack adequate NLP support. Guyanese Creole (Creolese) is the most widely spoken language in Guyana, yet it's a low-resource language due to limited written materials and formal recognition. This paper addresses this gap by creating a dedicated corpus and exploring machine translation capabilities for Guyanese Creole, aiming to foster its formal adoption and accessibility.
Methods
The core of this work involves the creation of the GuyLingo corpus. This was achieved through a multi-pronged approach:
- Expert Collaboration: Partnering with the University of Guyana, researchers curated, digitized, and manually transcribed original Guyanese Creole sources, including historical proverbs and contemporary texts like a COVID-19 children's book. Native Creole experts also manually created a corpus of common sayings.
- Online Resource Scraping: Web-based sources like language forums and blogs were scraped to capture colloquialisms, everyday conversations, and idiomatic expressions. These were then cleaned and verified.
To facilitate machine translation, the authors developed the Guyanese Creole Translation Tool. This web-based application uses GPT-4 as a generator, allowing Creole experts to iteratively create and refine English-Creole translation pairs. The tool incorporates advanced prompts with example translations for better generation quality.
For evaluating machine translation models, they fine-tuned standard NLP models like T5, BART, and Pegasus on their curated dataset.
Datasets
The primary dataset introduced is GuyLingo, a corpus of Guyanese Creole. It contains:
- 2373 Guyanese Creole sentences.
- A vocabulary size of 4177 unique Creole words.
- A translation dataset of 1969 English-Creole translation pairs, with 302 manually curated pairs used for testing.
The corpus draws from various sources, including:
- Historical texts (e.g., "The Proverbs of British Guiana" by Speirs, 1902).
- Contemporary materials (e.g., a COVID-19 children's book by Helen Patuck, 2020).
- Online resources (forums, blogs).
- Manually created common sayings by native experts.
Key Insights and Ideas
- The Power of Curated Corpora: The creation of GuyLingo highlights how dedicated, high-quality corpora are essential for advancing NLP for low-resource languages, enabling the development of sophisticated models.
- Bridging the Formal-Informal Divide: By developing tools and datasets for Guyanese Creole, the research aims to make the language more accessible for formal contexts, potentially leading to its official recognition.
- Challenges in English-Creole Translation: The paper identifies specific challenges in English-to-Creole translation, such as the lack of direct English equivalents for many unique Creole words and phrases, and inconsistencies in writing systems.
- GPT-4's Adaptability: The study demonstrates GPT-4's strong performance in Creole-English translation, showcasing its ability to adapt to new language pairs with minimal examples through in-context learning.
Key Results
- Machine Translation Performance:
- For English-to-Creole translation, the BART-Large model achieved the best performance among fine-tuned models, with a BLEU score of 12.11.
- For Creole-to-English translation, GPT-4 (Few-shot) achieved the highest BLEU score of 30.24, outperforming other models.
- GuyLingo Translation Tool: The tool, powered by GPT-4, successfully generated initial translations for the corpus, which were then manually reviewed and edited.
- Conversational AI: A demonstration of a conversational AI assistant (IRIS) speaking Guyanese Creole, fueled by GuyLingo, showcases the practical applications of the corpus.
Limitations and Future Work
- Limited Representation: The current GuyLingo corpus primarily focuses on Guyanese Creole. The authors acknowledge the need to include other indigenous languages of Guyana and additional Creole languages and dialects from the wider Caribbean for a more comprehensive dataset.
- Generalizability: Findings regarding the formal adoption of Creole languages may have limited generalizability to other regions or linguistic contexts.
- Language Evolution: Creole languages are dynamic. The static nature of the current corpus and models may not fully capture evolving linguistic usage, necessitating regular updates.
- Future Work: The authors suggest prioritizing the inclusion of other languages and dialects to ensure a more representative dataset and advocate for continued research to improve NLP for underrepresented languages.
Summary updated