alfredkondoro/swahili-nlu-intent-evals

# SwahiliNLU **A Broad-Domain Intent and Slot-Filling Dataset with Pragmatic Tone Annotations for Low-Resource NLU** --- ## Overview SwahiliNLU is the first broad-domain, natively constructed intent recognition and slot-filling dataset for Swahili. It covers **10 domains**, **175 intent classes** and **105 slot types**, and introduces the first **pragmatic tone annotation** (imperative / polite / conversational) for any African language NLU dataset. | Statistic | Value | |---|---| | Total utterances | 3,633 | | Domains | 10 | | Unique intents | 175 | | Unique slot types | 105 | | Total slot annotations | 5,079 | | Utterances with ≥1 slot | 3,037 (83.6%) | | Tone categories | 3 | | Inter-annotator agreement (tone) | κ = 0.87 | | Language | Swahili (sw) | --- ## Dataset The dataset is permanently archived on Zenodo: > **DOI:** 10.5281/zenodo.20521261 It is organised into 10 CSV files, one per domain: | Domain | File | Utterances | |---|---|---| | Communication | `communication.csv` | 513 | | Device Settings | `device_settings.csv` | 414 | | Finance | `finance.csv` | 461 | | Food Ordering | `food_ordering.csv` | 433 | | Health & Lifestyle | `health_lifestyle.csv` | 291 | | Home Automation | `home_automation.csv` | 208 | | Media & Entertainment | `media_and_entertainment.csv` | 376 | | News & Information | `news_and_information.csv` | 215 | | Ride-Hailing & Transport | `ride_hailing_and_transport_services.csv` | 179 | | Travel & Accommodation | `travel_transport_and_accomodation.csv` | 546 | ### Column format Each CSV has four columns: ``` Intent, Utterance, Slots, Label request_ride, Agiza Bolt kutoka Goba hadi Kariakoo, {"destination":"Kariakoo","pickup_location":"Goba","app_name":"Bolt"}, imperative ``` | Column | Description | |---|---| | `Intent` | Intent class label | | `Utterance` | Native Swahili utterance | | `Slots` | JSON dictionary of slot name → text span | | `Label` | Tone label: `imperative`, `polite` or `conversational` | --- ## Ton …