Swahili NLP Data Annotation: What Makes It Hard and How to Get It Right

Why Swahili Breaks Generic NLP Annotation Pipelines

Swahili (Kiswahili) is the most widely spoken African language, with an estimated 200 million speakers across East and Central Africa. It holds official status in Tanzania, Kenya, Uganda, and the Democratic Republic of Congo, and is the working language of the East African Community (EAC). Tanzania considers Swahili its primary national language; Kenya uses both Swahili and English officially; Swahili is growing as a lingua franca in South Sudan, Mozambique, and parts of the Central African Republic.

Swahili occupies an unusual position in NLP: it is simultaneously a high-speaker-count, strategically important language and one of the most under-resourced languages for AI annotation tooling. Most published Swahili NLP benchmarks are trained on formal Tanzanian Swahili news text, which represents a fraction of the digital Swahili corpus that annotation projects actually encounter. Kenyan urban Swahili — particularly Sheng — is structurally so different from formal Tanzanian Swahili that models trained on one perform near-randomly on the other. The practical consequence is that generic annotation pipelines produce drastically different quality on different sub-types of the Swahili corpus, with teams often discovering quality problems only after significant dataset investment.

The root causes are structural. Swahili's Bantu grammar creates agreement dependencies throughout sentences that non-native annotators systematically miss; the agglutinative verb morphology packs semantic content into complex word-internal structures that require native competence to parse correctly; and Sheng's rapid vocabulary evolution means annotation guidelines written six months ago may already be incomplete for new Sheng terms. Understanding each structural challenge in detail is the foundation for annotation pipeline design that actually works.

The Four Structural Challenges of Swahili NLP Annotation

1. The Bantu noun class system and agreement chains

Swahili's grammatical system is organised around approximately 15 noun classes, each identified by a prefix on the noun. Every noun belongs to a class, and that class membership triggers corresponding agreement prefixes on every associated verb, adjective, demonstrative, possessive, and relative pronoun in the same sentence. This agreement system creates what linguists call agreement chains — long strings of co-referential prefixes throughout a sentence that encode grammatical relationships.

The most annotation-relevant classes are: Class 1/2 (M-/Wa-, humans — mtu/watu, person/people), Class 3/4 (M-/Mi-, trees and long objects — mti/miti, tree/trees), Class 5/6 (Ji-/Ma-, augmentatives and some collectives — jiwe/mawe, stone/stones), Class 7/8 (Ki-/Vi-, tools and some abstracts — kitu/vitu, thing/things), Class 9/10 (N-/N-, most loanwords and animals — njia/njia, path/paths). Loanwords from English and Arabic almost universally enter Class 9 (N-class), giving them noun class membership that then drives agreement throughout the sentence.

For NER annotation, the agreement system creates two critical challenges. First, entity mentions may appear with class prefixes that visually change the entity token: the organisation 'Kampuni ya Safaricom' (Safaricom Company — Class 9/10 agreement) appears in different agreement contexts than 'Kikosi cha Safaricom' (Safaricom team — Class 7/8 agreement), but both refer to the same entity with different framing nouns. Second, coreference resolution requires tracking agreement chains across multiple clauses — the prefix on a verb three clauses later may co-refer to an entity named in clause one, which is the key structural connection for named entity coreference tasks.

For sentiment annotation, the agreement system changes the surface form of sentiment-bearing adjectives completely. The positive adjective stem -zuri (good/beautiful/well) appears as mzuri (Class 1/3), mzuri (Class 3), jizuri/zuri (Class 5/6), kizuri/vizuri (Class 7/8), nzuri (Class 9/10), uzuri (Class 11). Non-native annotators who do not recognise kizuri or vizuri as forms of 'good' miss 30–40% of positive sentiment markers in texts with Class 7/8 noun subject agreement.

2. Agglutinative verb morphology and intent encoding

Swahili verbs are highly agglutinative: a single verb word encodes subject agreement (noun class of the subject), tense/aspect marker, object agreement (noun class of the object), the verb root, and a final vowel that encodes mood (declarative, subjunctive, infinitive). The full structure is: SUBJECT AGREEMENT — NEGATION (optional) — TENSE/ASPECT — OBJECT AGREEMENT (optional) — VERB ROOT — FINAL VOWEL — EXTENSIONS (optional).

The annotation-relevant consequence is that a single Swahili verb word can encode the complete intent of a customer service message. The word sijalipa — unpacked: si- (1sg subject) + ja- (not-yet tense) + lip- (pay root) + a (declarative final vowel) — means 'I haven't paid yet,' a complete intent-laden statement. Nilitumia — ni- (1sg) + li- (past tense) + tum- (use/send root) + i- (applicative extension) + a (declarative) — means 'I sent (to someone),' a transaction confirmation. Hakulipa — ha- (negation) + ku- (3sg past) + lip- (pay) + a — means 'they didn't pay,' a third-party failure report.

For intent annotation in fintech and e-commerce contexts — where a large fraction of customer service messages are one or two Swahili verb words — non-native annotators who cannot parse the morphological structure classify by the verb root alone, missing the tense, negation, and object information that encode the actual intent. Research on Kenyan M-Pesa customer service data found that morphological parsing errors produced intent misclassification rates of 35–45% on single-verb-word messages — which comprised approximately 20% of the corpus. These errors are structurally invisible to annotators who do not have native Swahili morphological competence.

3. Sheng: Nairobi's urban code-mix

Sheng is a contact language that emerged in Nairobi's Eastlands communities in the 1970s and has become the dominant informal register among Kenyan urban youth. It is not a dialect of Swahili — it is a code-mix that uses Swahili grammatical structure as a matrix language while freely substituting vocabulary from Kikuyu, Luo, Luhya, and English, with continuous lexical innovation driven by community identity.

Sheng's defining properties create annotation challenges at every level. Vocabulary is highly dynamic — new Sheng terms are coined and spread rapidly through social media, and a term accepted in one neighbourhood may be unknown or have different connotations in another. English words are morphologically integrated: the English verb 'call' is integrated as Swahili Class 9 infinitive kukol (to call); the English noun 'job' is integrated as Swahili noun kazi/jobu depending on the Sheng variant and community. Kikuyu vocabulary inserts fluidly: thao (thousand, from Kikuyu) replaces standard Swahili elfu in Sheng financial communication. Research from the University of Nairobi and Oxford on Kenyan social media corpora (2022) estimates 55–70% of Nairobi-based Twitter and WhatsApp text uses Sheng vocabulary or code-mixing conventions.

For annotation pipelines targeting Kenyan digital products — M-Pesa, Safaricom services, Jumia Kenya, e-commerce platforms, and Kenyan government digital services — Sheng is not an edge case. It is the dominant register of the target user population. Annotation pipelines calibrated on formal Tanzanian Swahili news text produce error rates of 40–55% on Sheng-heavy Kenyan corpora, because Sheng vocabulary is treated as out-of-vocabulary noise and the noun class agreement chains on Sheng-integrated English verbs and nouns are parsed incorrectly. The result is datasets that systematically mislabel the intent and sentiment of the largest segment of the Kenyan digital user base.

4. Kenyan vs Tanzanian register divergence

Tanzania's Swahili is considered the standard — Zanzibar and coastal Tanzania are the historical home of Swahili, and formal Tanzanian Swahili is the basis of most published Swahili NLP benchmarks. Kenyan Swahili is a second-language or lingua-franca Swahili for many speakers, heavily influenced by English and the major Kenyan Bantu and Nilotic languages. The result is two distinct registers that diverge in vocabulary, pragmatic conventions, and the extent of code-mixing with English.

For annotation projects, the register divergence creates two distinct quality problems. First, annotation guidelines written for Tanzanian formal Swahili do not cover Kenyan register conventions — politeness markers, customer-service discourse patterns, and domain-specific vocabulary are different. A Kenyan M-Pesa customer writing Naskia vibaya sana, naomba msaada (I feel very bad, I request help) is expressing distress and an urgent service escalation; a Tanzanian speaker using the same sentence in a different context may be expressing mild dissatisfaction. Intent classification calibrated on Tanzanian guideline examples will misclassify Kenyan urgency signals at rates of 20–30%.

Second, the large English code-switching in Kenyan Swahili creates annotation complexity that formal Tanzanian corpora do not prepare annotators for. A Kenyan customer service message that switches fluidly between Swahili, Sheng, and English in a single sentence — Nimejaribu kukucontact several times lakini connection inashindwa, sasa balance yangu iko wapi? (I've tried to contact you several times but the connection fails, now where is my balance?) — requires annotation competence in all three registers simultaneously. Annotators trained on formal Tanzanian Swahili alone cannot reliably parse, segment, or intent-classify this kind of message.

The Swahili NLP Tooling Ecosystem

Swahili NLP has a smaller but growing open-source tooling ecosystem, anchored by the Helsinki NLP group, the Masakhane African NLP research community, and Lelapa AI (a South Africa-based AI lab focused on African language models).

Case Study: Kenyan Fintech — Customer Intent Classification Recovery

In late 2025, a major Kenyan mobile money and financial services platform (comparable in scale to M-Pesa or Airtel Money Kenya) needed 42,000 annotated customer service messages in Swahili for an intent classification model to route customer queries automatically and reduce contact centre volume. Messages were drawn from in-app chat and SMS, including a mix of formal Swahili, informal Kenyan Swahili, Sheng, and English code-switching. Approximately 30% of messages were single-verb or two-token messages — a single agglutinative verb word constituting the complete customer message.

The initial annotation run used a multilingual crowdsourcing platform with Swahili-certified annotators. After 10,000 records, the internal NLP team reviewed a validation sample against a 350-record gold standard annotated by senior Kenyan linguists and found:

• Intent accuracy of 58.3% against the gold standard — against a 90% project target
• Single-verb-word messages had intent accuracy of 41.7% — crowd annotators were classifying by verb root without parsing tense, aspect, negation, and object agreement
• 52% of Sheng messages were mislabelled — Sheng vocabulary was treated as out-of-vocabulary and messages were classified as noise or assigned default intents
• Noun class agreement mismatch on sentiment adjectives produced 38% sentiment misclassification on messages containing Class 7/8 agreement forms of sentiment-bearing adjectives
• Kenyan urgency markers ('haraka, haraka,' 'tafadhali sana,' 'sio sawa kabisa') were not recognised as escalation signals, causing 29% misclassification of urgent escalation messages as routine enquiries

The team rebuilt the annotation pipeline with native Swahili-speaking annotators and proper pre-processing:

• Helsinki NLP Swahili morphological decomposition on all 42,000 records, displaying verb morpheme breakdown alongside raw text for annotators
• Register classification to identify Sheng-dominant, formal Swahili, and code-mixed messages — routing 40% Sheng-dominant messages to annotators with Nairobi urban youth register qualification
• Kenyan Swahili dialect routing for messages with Kenyan-specific pragmatic markers, vocabulary, and urgency conventions — explicitly excluding Tanzanian-only calibrated annotators from these records
• Annotation guideline additions: noun class agreement form tables for the 20 most common sentiment adjectives, Kenyan urgency marker intent attribution rules, Sheng lexicon reference for the 150 most frequent Sheng terms in financial communication, and English code-switching domain-identification policy
• Double annotation on 15% of records with kappa measurement per intent class and per register stratum

Results on the re-annotated corpus:

The native-speaker annotation cost was 3.5× higher per record. The 10,000 crowd-annotated records — particularly the Sheng and single-verb-word subsets — were structurally unusable and discarded. The downstream intent model, trained on native-speaker data, achieved 90.1% accuracy on production holdout, enabling the platform to automatically route 79% of customer service queries with a false-routing rate of 3.4% — well below the platform's 5% escalation target — and releasing contact centre capacity for complex financial escalations and fraud case handling.

Swahili Annotation Guidelines: What Generic Templates Miss

Swahili annotation guidelines adapted from English or generic multilingual templates routinely omit the language-specific instructions that prevent the most common errors. Critical inclusions are:

• Noun class agreement form tables: For each sentiment-bearing adjective and intensifier in the guideline, provide the complete agreement form table across all 15 noun classes. Include -zuri (good), -baya (bad), -nzuri (fine), -kubwa (big/great), -dogo (small/little), -sawa (okay/correct), and the most domain-relevant sentiment vocabulary for the project. This table is the single highest-value addition for sentiment annotation projects.
• Verb morphology intent attribution table: Provide a morphological parsing guide for the 20 most common Swahili intent-encoding verb forms in the target domain. For fintech: sijalipa (I haven't paid yet — not-yet tense, payment escalation), nililipa (I paid — past tense, transaction confirmation), hakulipa (they didn't pay — negation + past + 3sg, third-party failure report), nitalipia (I will pay — future, payment intent). Pair each form with the correct intent label.
• Sheng lexicon reference: Provide a reference table of the 100–200 most frequent Sheng terms for the project domain with their standard Swahili equivalents and any sentiment/pragmatic implications. Specify the update cadence — Sheng vocabulary evolves; guideline Sheng lexicons should be reviewed every three months for active annotation projects.
• Kenyan vs Tanzanian register routing: Explicitly define the criteria for routing messages to Kenyan-calibrated vs Tanzanian-calibrated annotators. Signal words: Sheng vocabulary, Kenyan-specific organisation names (M-Pesa, Safaricom, Equity Bank), Kenyan urgency markers, and English code-switching above a threshold. Tanzanian routing: absence of Sheng, formal Swahili syntax, Tanzanian-specific organisation names (CRDB Bank, Vodacom Tanzania, NMB Bank).
• Single-verb message policy: Define explicitly that single-verb-word messages must be morphologically decomposed before intent classification. Provide a morphological decomposition format for annotators — either a displayed decomposition or a reference parsing table — and specify that intent is to be assigned from the full morphological meaning, not from the verb root alone.

The East African AI Market and Swahili Annotation Demand

East Africa's combined GDP across Kenya, Tanzania, Uganda, and DRC exceeds $450 billion (2024 World Bank data), and the region's technology sector is growing at 15–20% annually. Kenya's technology ecosystem — anchored by Safaricom's M-Pesa platform (50+ million active users, processing $314 billion in annual transactions as of 2024), the Nairobi-based startup ecosystem (iProcure, M-Kopa, Twiga Foods, Copia), and the growing enterprise AI sector — represents the largest single source of Swahili NLP annotation demand globally.

Tanzania's government digitisation programme and the EAC Digital Transformation Strategy are directing investment toward Swahili-language AI for regional government services, including the EAC Single Customs Territory, the East African Monetary Union preparations, and the Pan-African Payment and Settlement System (PAPSS). These government programmes require formal Tanzanian Swahili annotation at production scale. International demand comes from the World Bank, UNICEF, and WHO — all of which have significant Swahili-language AI requirements for health AI, agricultural extension AI, and humanitarian response systems in East Africa.

Our multilingual annotation services include native Swahili-speaking annotators for NER, sentiment, intent, and text classification tasks — with Sheng normalisation, noun class agreement parsing, and Kenyan/Tanzanian dialect routing included as standard. For teams building Swahili annotation alongside other African or multilingual requirements, our native-speaker annotation network covers Swahili alongside Arabic, French (African varieties), and 120+ other languages with consistent quality standards. For broader NLP annotation capabilities, our text annotation services provide end-to-end NLP annotation including NER, sentiment analysis, intent classification, and coreference resolution across Swahili and all major African languages.

Related Reading

If you are building Swahili NLP annotation alongside other language or multilingual requirements, these posts cover annotation challenges and strategies for related contexts:

• How Does Multilingual Annotation and Localization Work for Global AI? — Multi-language annotation workflows when Swahili is one of several target languages in an African or global product rollout
• How Much Does Using Native-Speaker Annotators Improve Multilingual AI? — Quantifying the quality lift from native-speaker annotation across language families, with case studies relevant to African languages
• How Is Multilingual Speech Transcription Annotation Done at Scale? — ASR annotation workflows for Swahili and multilingual African audio datasets, including accent and dialect handling

Frequently Asked Questions