Compound Verbs in SwahiliEnglish Machine Translation

Swahili (Kiswahili) is an agglutinative Bantu language in which a single verb can encode subject, object, tense, aspect, mood, polarity and several derivational elements. One of the most challenging phenomena for statistical and neural machine translation (SMT/NMT) systems is the compound verb construction, where a verb stem combines with a series of prefixes and suffixes to express what in English would require a multiword phrase.

What Are Compound Verbs?

In Swahili a verb complex consists of:

• Subject prefix (SP): indicates the subject noun class.
• Tenseaspectmood (TAM) marker: e.g., -li- (past), -na- (present progressive), -ta- (future).
• Object prefix (OP): marks the object noun class.
• Verb root (VR): the lexical core.
• Derivational suffixes: causative -isha, applicative -ia, reversive -ua, etc.
• Final vowel: usually -a (indicative) or -i (imperative).

Example:

nitakupendawanaa

Breakdown:

• ni = I (subject prefix, class 1)
• ta = future
• ku = you (object prefix, class 2)
• pend = love (root)
• -a = indicative
• -wa = reciprocal (you love each other)
• -naa = progressive aspect + present marker

English rendering: I will love you (mutually) now.

Why Compound Verbs Challenge MT

1. Data sparsity The same semantic idea can be expressed by dozens of different morphological combinations. In a parallel corpus many of these forms appear only a handful of times, making reliable lexical alignments difficult.
2. Segmentation ambiguity Deciding where a verb ends and the next word begins is nontrivial. Wordtokenizers that simply split on whitespace will treat the whole complex as a single token, starving the model of subword evidence.
3. Longdistance dependencies The object prefix may refer to a noun that appears several clauses earlier, and the causative or applicative suffixes alter the argument structure. Traditional phrasebased MT struggles to capture these reordering patterns.
4. Morphological overgeneration NMT models using bytepair encoding (BPE) may generate impossible suffix sequences (e.g., -ishaawa) because the subword vocabulary does not encode the grammatical ordering constraints.

Approaches to Handling Compound Verbs

1. Morphological Preprocessing

Segment the verb complex into its morphemes before feeding it to the translation model.

nitakupendawanaa  [ni] [ta] [ku] [pend] [a] [wa] [na] [a]

Benefits:

• Reduces vocabulary size.
• Allows the model to learn systematic correspondences (e.g., -isha make .)

2. Factored Neural Translation

Encode each morpheme as a separate factor (e.g., stem, TAM, object). The decoder then combines the predicted factors to generate a fluent English phrase.

3. Joint MorphologyTranslation Models

Train a single network that simultaneously performs morphological analysis (segmentation + tagging) and translation. The shared encoder learns representations that respect Bantu morphotactics.

4. Postediting Rules

After an NMT pass, a rulebased component rewrites literal translations of morphemes into idiomatic English.

I am loveyoureciprocalpresent  We love each other.

Case Study: Neural Model with Morphological Segmentation

Researchers at Masakhane trained a Transformer on a 1millionsentence SwahiliEnglish corpus where each verb was segmented using the kwanza analyzer. The results:

• BLEU improvement of +3.4 over a baseline that kept verbs unsplit.
• Reduction in UNK tokens from 12% to 4%.
• Qualitative analysis showed more accurate handling of causative -isha (make/let) and applicative -ia (for/at).

Practical Guidelines for Developers

1. Select a reliable morphological analyzer. Tools such as AFR, KuSwahiliParser, or the opensource morphology-swahili library provide tokenlevel segmentation with POS tags.
2. Maintain alignment between morpheme tags and English glosses. Create a small parallel gloss dataset (e.g., nitapikaa I will cook) to finetune the decoder.
3. Use subword units that respect morpheme boundaries. Apply BPE after segmentation, not before.
4. Incorporate linguistic features. Encode nounclass information as an additional embedding; this helps the model decide when an object prefix should be rendered as him/her/it.
5. Evaluate with targeted test sets. Construct sentences that isolate each derivational suffix (causative, applicative, reciprocal) to verify that the system learns the correct mapping.

Future Directions

• Zeroresource transfer. Leverage highresource Bantu languages (e.g., Zulu) with similar verb morphology to improve Swahili MT without extra parallel data.
• Multitask pretraining. Combine language modeling, morphological tagging, and translation in a single massive transformer.
• Interactive translation tools. Provide users with a morphemebymorpheme edit interface, allowing manual correction of rare or ambiguous compounds.

Conclusion

Compound verbs are the heart of Swahili grammar and a major obstacle for accurate machine translation into English. By explicitly modeling the internal structure of verb complexesthrough segmentation, factored representations, or joint morphologytranslation networkssystems can achieve substantial gains in fluency and adequacy. Continued collaboration between computational linguists, native speakers, and the broader opensource community will be essential to bridge the remaining gaps and produce robust SwahiliEnglish translation tools.