Congolese Swahili Machine Translation

Background

Swahili (Kiswahili) is one of the most widely spoken Bantu languages in Africa, with an estimated 100million speakers. In the Democratic Republic of the Congo (DRC), a distinct dialect known as Congolese Swahili (also called Kingwana) is spoken by several million people, especially in Katanga, Tanganyika and the eastern provinces. While the core grammar of Kingwana aligns with Standard Swahili, centuries of contact with Lingala, French, and local languages have produced lexical borrowings, phonological shifts, and syntactic variations that set it apart from the East African norm.

The rise of internet connectivity and mobile messaging in the DRC has created a growing demand for translation tools that can bridge Congolese Swahili with global languages such as English, French, and Portuguese. Traditional rulebased systems fail to capture the fluidity of contemporary usage, prompting researchers to explore statistical and neural machine translation (NMT) techniques.

Linguistic Features of Congolese Swahili

Lexical Borrowing

Kingwana incorporates a large number of loanwords from French (moto moto car), Lingala (mbongo money) and Portuguese (ponto point). These borrowings often retain their original phonology, creating pronunciation patterns that differ from Standard Swahili.

Verb Morphology

Swahili verbs are agglutinative: ni-na-ku-soma (I am reading you). In Kingwana, speakers sometimes simplify the subject prefix (e.g., ni n) or omit the object marker, especially in informal text messages. This leads to tokenization challenges for MT systems.

CodeSwitching

Urban speakers frequently switch between Kingwana, Lingala, and French within a single sentence (Nakutuma le document ya PDF). A robust translation model must recognize and appropriately handle mixed language segments.

Orthographic Variation

There is no single standardized orthography for Kingwana. Variants such as kazi versus kadzi for work coexist. Normalization pipelines are therefore a prerequisite for any largescale MT effort.

Data Resources

Highquality parallel corpora are the lifeblood of modern MT. For Congolese Swahili the following resources have proven most useful:

• CCAligned a collection of crawled web pages aligned at the sentence level, containing a modest amount of KingwanaEnglish pairs.
• OpenSubtitles subtitles from movies and series that include usergenerated Kingwana translations; after cleaning, they provide ~250k sentence pairs.
• SMS and WhatsApp corpora anonymized chat logs donated by NGOs working in the DRC; these capture informal codeswitching patterns.
• Religious texts the New Testament has been translated into Kingwana; while domainspecific, it offers wellaligned, highquality data.

Because the total volume of clean, aligned data is still below 1million sentence pairs, many projects resort to backtranslation and synthetic data generation to augment the training set.

Model Approaches

Statistical Machine Translation (SMT)

Early experiments used phrasebased SMT (Moses) with lexical smoothing to handle outofvocabulary (OOV) items. While SMT could capture basic word order, it struggled with the rich morphology and codeswitching patterns typical of Kingwana.

Neural Machine Translation (NMT)

Current stateoftheart systems rely on the Transformer architecture. Two strategies have emerged:

1. Multilingual Training joint training on Swahili (Standard), Lingala, French, and English enables the model to share subword vocabularies and transfer knowledge from highresource language pairs.
2. DomainAdaptive FineTuning a base model pretrained on large SwahiliEnglish corpora (e.g., JW300) is finetuned on the smaller Kingwana datasets, improving dialectspecific performance.

Subword Segmentation

BytePair Encoding (BPE) with 32k merges balances coverage of rare loanwords and preservation of morphological stems. Experiments show a 1.5BLEU gain when BPE is trained on a combined SwahiliKingwana corpus rather than on Standard Swahili alone.

Handling CodeSwitching

Tagaware models prepend language tags (<2en>, <2fr>) to each token, allowing the decoder to condition on the source language. Additionally, a lightweight languageidentification layer filters mixed sentences before translation.

Evaluation

Quantitative metrics (BLEU, ChrF) are complemented by human evaluation because automatic scores often mask errors in codeswitching handling. A typical evaluation protocol includes:

• BLEU on test set 27.8 for a multilingual Transformer finetuned on Kingwana data.
• ChrF better reflects characterlevel similarity, yielding 38.5.
• Human adequacy & fluency rated on a 5point scale by bilingual speakers; average adequacy 4.1, fluency 4.0.
• Codeswitching accuracy proportion of correctly translated borrowed words; achieved 82% after tagaware training.

Error analysis highlights three recurring issues:

1. Mistranslation of French loanwords that have no Swahili equivalent.
2. Overgeneration of subject prefixes, producing ungrammatical forms such as ni-na where Kingwana prefers n.
3. Inconsistent handling of punctuation in informal chat data.

Future Directions

Data Expansion

Communitydriven collection of parallel sentences, especially from educational and health NGOs, can quickly increase domain diversity. Crowdsourcing platforms that support Swahili dialects are being piloted in Kinshasa.

Interactive MT

Integrating MT into messaging apps with onthefly correction allows users to edit suggestions, feeding back into model finetuning in a continuallearning loop.

Dialectaware Pretraining

Large language models (LLMs) pretrained on multilingual African text can be adapted to Kingwana through instruction tuning, potentially reducing the need for massive parallel corpora.

Ethical and Social Considerations

Transparency about model limitations, especially regarding culturally sensitive content, is essential. Moreover, preserving linguistic diversity must be balanced against the risk of homogenizing dialects under a single standard Swahili model.

Key Takeaways

• Congolese Swahili is a vibrant, borrowingrich dialect that presents unique challenges for machine translation.
• Limited parallel data can be mitigated through multilingual training, backtranslation, and subword techniques.
• Handling codeswitching and orthographic variation requires languageaware preprocessing and tagging.
• Human evaluation remains crucial to capture the adequacy of translations in informal contexts.
• Future progress hinges on community data collection, interactive systems, and ethical deployment.