Introduction
Minangkabau (Baso Minang) is an Austronesian language spoken by over 7 million people, primarily in West Sumatra, Indonesia. Despite its significant speaker base, it remains a "low-resource" language in the digital sphere, as formal communication in the region is dominated by Bahasa Indonesia. This paper presents the first major effort to build Natural Language Processing (NLP) tools for Minangkabau, introducing foundational datasets and benchmarking performance across sentiment analysis and machine translation tasks.
Methodology and Data Construction
The researchers addressed the data scarcity by developing three primary resources: a bilingual dictionary, a sentiment corpus, and a parallel machine translation (MT) corpus.
Dictionary and Parallel Corpus Building
The team constructed a bilingual dictionary containing 11,905 word pairs by translating top terms from Minangkabau Wikipedia. For the MT task, they developed an automated pipeline to align sentences between Minangkabau and Indonesian Wikipedia articles. This process utilized ROUGE-1 scores and trigram filtering to ensure the quality of the resulting 16,371 parallel sentence pairs.
Sentiment Corpus
The authors released a sentiment corpus of 5,000 sentences (1,481 positive, 3,519 negative) manually translated from Indonesian social media and hotel reviews. This dataset captures the unique linguistic challenges of Minangkabau, such as mixed polarity and the use of negative words to convey positive sentiment.
Key Insights: The "Intelligibility" Trap
A core finding of the study is that linguistic proximity to a high-resource language does not guarantee successful "zero-shot" transfer. While Minangkabau and Indonesian share approximately 55% of their vocabulary, the researchers observed an 11.41-point drop in F1 score when applying an Indonesian sentiment model to Minangkabau text. This confirms that "close enough" is insufficient for high-quality NLP, necessitating dedicated local resources.
Furthermore, the study found that in low-resource settings, simpler statistical models can be more robust than complex neural networks. In zero-shot scenarios, Naive Bayes outperformed multilingual BERT (mBERT), suggesting that basic vocabulary overlaps are sometimes more reliable than embeddings that have not seen enough of the target language.
Experimental Results
- Sentiment Analysis: The best performance was achieved by mBERT when fully trained on Minangkabau data, reaching an F1-score of 75.91.
- Machine Translation: A simple Word-to-Word (W2W) approach using the new dictionary achieved the highest BLEU score (64.54), outperforming both LSTM (63.77) and Transformer (56.25) models. The neural models struggled with the small dataset size, often "hallucinating" irrelevant information.
Conclusion and Future Work
This research establishes a vital baseline for Minangkabau NLP but highlights the limitations of current neural approaches in low-resource contexts. Future work should focus on expanding the corpus to include colloquial text, addressing dialectal variations, and developing specialized training schemes to reduce model hallucinations in data-scarce environments.
Summary updated