Towards Computational Linguistics in Minangkabau Language: Studies on Sentiment Analysis and Machine Translation

The increasing linguistic diversity worldwide emphasizes the importance of digital preservation for minority languages like Minangkabau. Historically, research focused on phonetics, morphology, and lexicography (Rusmali et al., 1985; Crouch, 2009), but lacked systematic corpora for NLP applications. With social media’s rise, Minangkabau’s online presence grew, yet the scarcity of annotated datasets hindered computational efforts. Similar low-resource languages such as Javanese and Sundanese have seen progress through speech chain models and translation systems (Wibawa et al., 2018; Suryani et al., 2016), but Minangkabau remains underexplored. This study aims to fill this gap by constructing the first parallel corpus, evaluating classical and neural models, and exploring transfer learning strategies, thus laying a foundation for future NLP applications and cultural preservation efforts.

The core challenge lies in the severe lack of annotated corpora, standardized resources, and dialectal diversity for Minangkabau. Existing linguistic resources are limited to dictionaries and phonological descriptions, insufficient for training data-hungry models like deep neural networks. The dialectal variation complicates corpus collection and model generalization, while the limited vocabulary coverage hampers transfer learning. Additionally, the absence of large-scale parallel corpora restricts the development of high-performance translation and classification systems. Overcoming these bottlenecks is crucial for enabling effective NLP tools that can support language revitalization, digital archiving, and cultural promotion, especially in a context where the language is primarily oral and informal.

This work introduces several innovations: 1) manual construction of a 16K sentence parallel corpus for Minangkabau–Indonesian translation, filling a critical resource gap; 2) development of a bilingual dictionary with 11,905 entries, leveraging Wikipedia data; 3) systematic comparison of traditional ML models and deep Seq2Seq architectures in low-resource settings; 4) pioneering application of MBERT for indigenous language classification. The integration of lexicon-based translation with neural models demonstrates a practical approach to low-resource NLP, emphasizing the synergy between linguistic resources and machine learning. The research also pioneers the evaluation of cross-lingual transfer in a minority language context, providing new insights into model robustness and resource efficiency.

• �� Data collection: Extracted Wikipedia articles and social media posts in Minangkabau, manually translated top words, and aligned sentences using ROUGE-1 scores. • Dictionary creation: Selected top 20,000 words from Wikipedia, manually translated into Indonesian, filtered via KBBI, resulting in 11,905 entries. • Corpus construction: Aligned 111,430 document pairs, segmented into sentences, and filtered to 345,146 sentence pairs with ROUGE-1 > 0.5. • Model training: Employed classical ML (Naive Bayes, SVM, Logistic Regression) with byte-pair encoding, deep models (Bi-LSTM, Transformer), and MBERT, fine-tuning each on Indonesian data. • Evaluation: Used BLEU scores for translation, F1-score for sentiment analysis, with 5-fold cross-validation and manual validation for corpus quality. • Experiments: Conducted zero-shot transfer tests, comparing performance on Minangkabau and Indonesian translations, analyzing vocabulary overlap and out-of-vocabulary issues.

The experiments involved training models on Indonesian data and testing on Minangkabau, assessing zero-shot transfer. For sentiment analysis, datasets from Koto and Rahmaningtyas (2017) were translated into Minangkabau, with 1481 positive and 3519 negative labels. Models included Naive Bayes, SVM, Logistic Regression, Bi-LSTM with pre-trained fastText embeddings, and MBERT. For machine translation, models trained on Wikipedia and Sentiment corpora used Word-to-Word, LSTM, and Transformer architectures, evaluated via BLEU scores. The translation experiments revealed that dictionary-based W2W translation achieved BLEU scores over 60, outperforming neural models, especially on out-of-domain data. The sentiment models showed that simple models like Naive Bayes could outperform deep models in zero-shot scenarios, emphasizing vocabulary overlap's importance. Manual evaluation confirmed the effectiveness of dictionary-assisted translation and the challenges posed by dialectal variation.

The key findings include: 1) Naive Bayes achieved 68.49 F1 in Minangkabau sentiment classification, surpassing MBERT; 2) Word-to-Word translation with a bilingual dictionary reached BLEU scores of 64.54 (MIN→ID), outperforming LSTM (56.25) and Transformer (43.50); 3) Cross-lingual transfer performance significantly drops in out-of-domain datasets, highlighting vocabulary and dialect limitations; 4) Manual validation confirmed dictionary-assisted translation's high adequacy and fluency; 5) The high vocabulary overlap (~75%) facilitated transfer learning, but out-of-vocabulary issues remain a bottleneck.

The developed corpora and models enable practical applications such as sentiment monitoring on social media, automated translation for cultural content, and digital archiving of Minangkabau texts. These tools can support language revitalization, content localization, and cross-cultural communication. The approach can be extended to other low-resource languages, especially in multilingual societies, fostering linguistic diversity and cultural preservation. In industry, such models can be integrated into chatbots, virtual assistants, and translation services, providing accessible language technology for indigenous communities. Long-term, these efforts could lead to comprehensive NLP ecosystems supporting minority language speakers and cultural heritage.

The corpus size remains limited, with manual translation introducing noise and bias. Deep neural models underperform in out-of-vocabulary scenarios, especially on domain-specific or dialectal data. The models' robustness across dialects and informal registers needs further validation. Computational costs for training large models are high, and the current approach relies heavily on lexicon resources, which may not scale well. Future work should focus on expanding data, automating annotation, and developing more robust transfer strategies to address these challenges.