[Paper Review] Towards a parallel corpus of Portuguese and the Bantu language Emakhuwa of Mozambique

Felermino D. M. A. Ali, Andrew Caines+1 more|arXiv (Cornell University)|Apr 12, 2021

Natural Language Processing Techniques11 references4 citations

TL;DR

This paper presents the creation of a 47,415 sentence pair parallel corpus of Portuguese and Emakhuwa, a low-resource Bantu language spoken in Mozambique, sourced from religious, legal, and educational texts. The corpus, which contains 699,976 Emakhuwa and 877,595 Portuguese word tokens, is being normalized for research use and includes a baseline neural machine translation model achieving 12.42 BLEU on Portuguese-to-Emakhuwa translation.

ABSTRACT

Major advancement in the performance of machine translation models has been made possible in part thanks to the availability of large-scale parallel corpora. But for most languages in the world, the existence of such corpora is rare. Emakhuwa, a language spoken in Mozambique, is like most African languages low-resource in NLP terms. It lacks both computational and linguistic resources and, to the best of our knowledge, few parallel corpora including Emakhuwa already exist. In this paper we describe the creation of the Emakhuwa-Portuguese parallel corpus, which is a collection of texts from the Jehovah's Witness website and a variety of other sources including the African Story Book website, the Universal Declaration of Human Rights and Mozambican legal documents. The dataset contains 47,415 sentence pairs, amounting to 699,976 word tokens of Emakhuwa and 877,595 word tokens in Portuguese. After normalization processes which remain to be completed, the corpus will be made freely available for research use.

Motivation & Objective

• To address the scarcity of parallel corpora for low-resource African languages by creating a large-scale, multilingual dataset for Portuguese and Emakhuwa.
• To support the development of neural machine translation (NMT) systems for Emakhuwa, a Bantu language with limited computational and linguistic resources.
• To improve accessibility and research potential for African language NLP by compiling and normalizing diverse text sources including religious, legal, and educational materials.
• To establish a foundation for future expansion of the corpus with diverse domains and improved linguistic normalization.
• To provide a publicly available, standardized dataset to support reproducible research and model benchmarking in low-resource translation settings.

Proposed method

• Collected parallel texts from multiple sources: the Jehovah’s Witness website (JW), African Story Book (ASB), Centro Catequético Paulo VI (CCA), Universal Declaration of Human Rights (HR), and Mozambican Land Law (LL).
• Performed OCR processing on scanned documents, followed by manual correction of common OCR errors such as 'ì' misrecognized as 'T' and 'á' as '6'.
• Conducted sentence-level segmentation and filtering to remove incorrectly recognized or malformed sentences.
• Executed manual alignment of Portuguese and Emakhuwa sentences to ensure accurate parallel sentence pairs.
• Addressed spelling inconsistencies, including variations in the use of 'j' vs 'c' and 'erya' vs 'jerya' in verb forms, and 'biblia' vs 'biibiliya' for 'bible'.
• Trained a baseline NMT model using OpenNMT-py with a 2-layer LSTM encoder and decoder (500 hidden units), trained on a 90:10 random split of the data.

Experimental results

Research questions

• RQ1 How can a large-scale, high-quality parallel corpus be constructed for a low-resource African language like Emakhuwa?
• RQ2 What are the main challenges in collecting and normalizing parallel text for Emakhuwa, particularly regarding spelling variation and OCR errors?
• RQ3 What baseline performance can be achieved using a standard NMT architecture on the newly created Emakhuwa-Portuguese parallel corpus?
• RQ4 How can diverse text sources—religious, legal, educational—be integrated into a single, usable parallel corpus for low-resource language translation?
• RQ5 What role can community-driven initiatives and open data sharing play in advancing NLP for under-resourced African languages?

Key findings

• The final corpus comprises 47,415 sentence pairs, with 699,976 word tokens in Emakhuwa and 877,595 word tokens in Portuguese.
• The dataset is derived from 125 document pairs across five distinct sources, including religious texts (JW, CCA), educational content (ASB), legal documents (LL), and human rights declarations (HR).
• A baseline NMT model achieved a BLEU score of 12.42 on Portuguese-to-Emakhuwa translation and 13.09 on Emakhuwa-to-Portuguese translation.
• Significant spelling inconsistencies were identified, such as 'cerya' vs 'jerya' in verb forms and 'biblia' vs 'biibiliya' for 'bible', necessitating future normalization.
• OCR errors, including misrecognition of 'ì' as 'T' and 'á' as '6', required manual correction during preprocessing.
• The corpus is expected to be released publicly upon completion of normalization, with plans to expand it using additional data sources and domain diversity.