AfriSpeech/afro-topics

AfroTopics

A topic-classification layer over AfriSpeech/africa-corpus (Bible text) and AfriSpeech/african-corpus-jw (JW.org publications), classifying the English anchor text into a fixed topic taxonomy with Gemini, then propagating the label to every aligned African-language translation via the shared verse_key.

Taxonomy

• agriculture
• health
• governance
• economy
• education
• technology
• infrastructure
• environment
• culture
• security

Every item is categorized into one of the 10 topics above — there is no none/catch-all option. An earlier pass allowed a none label for content with no clear real-world subject matter (narrative, genealogy, poetry, fragments, captions); on review that made the dataset less useful, so those items were re-classified with none removed from the label set, forcing a best-fit real-topic decision even where the fit is loose (e.g. Genesis 1's creation narrative → environment). Any remaining un-reclassified items were filtered out, ensuring that 100% of items in the published dataset carry a concrete topic label.

Structure

• bible/english_bible.csv — verse_key, eng, topic (28711 verses)
• bible/{iso1}_{iso3}.csv — verse_key, local, topic, one file per high-resource Bible-corpus language (23 languages)
• jw/english_jw.csv — verse_key, eng, topic (386427 paragraphs)
• jw/{Name}_{code}.csv — verse_key, local, topic, one file per JW-corpus African language (135 languages)

The Bible and JW portions use different key namespaces (BOOK.chapter.verse vs docid#pid) and are kept as separate files rather than merged.

Method

1. Pulled the English anchor text (english_cache.csv) from each source dataset.
2. For the Bible corpus, classified verses with Gemini (gemini-2.5-flash).
3. For the JW corpus, first restricted to paragraphs that have at least one African-language parallel translation anywhere in the corpus (paragraphs with zero parallels can't be used for a parallel dataset anyway), then classified those with Gemini (gemini-2.5-flash-lite, thinking disabled).
4. Second pass: items labeled none in step 2/3 were re-classified with none removed from the label set, forcing the model to pick whichever of the 10 real topics fits best. Items not re-classified were omitted so all retained examples carry a valid topical label.
5. Joined the resulting verse_key -> topic labels onto each high-resource language's per-language file.
6. For Bible-corpus languages with multiple available translation versions, picked the single largest/most-complete version (preferring standard script over alternate-script variants, e.g. Latin-script Hausa over the Arabic-script edition).

Caveats

• Labels are LLM judgments on the English text only, not human-validated, and not verified against the African-language text itself — treat them as weak/silver labels. A meaningful fraction were forced best-fit choices on content with no strong real-world topical grounding (narrative, genealogy, fragments) — see the taxonomy note above.
• An earlier attempt classified this data with a pure sentence-embedding nearest-neighbor approach (BAAI/bge-large-en-v1.5 against category description embeddings); it was abandoned after inspection showed low, non-discriminating confidence scores and frequent misclassification of fragments/captions. Gemini was used instead for both corpora.
• Category distribution is uneven and reflects the source material: Bible narrative skews toward culture and security (religious/war content); JW literature covers the other topics more evenly but still skews toward culture.

Source & license

Bible text is derived from public Bible translations via YouVersion (see AfriSpeech/africa-corpus for terms). JW text is an alignment index over publicly accessible jw.org pages, content (c) Watch Tower Bible and Tract Society of Pennsylvania (see AfriSpeech/african-corpus-jw for terms). This repository's own contribution (the topic labels and the joined structure) is released under the same terms as the two source datasets.