Researchers have developed CamSpokenNLI, a new pilot dataset designed to evaluate cross-variety transfer in Natural Language Inference (NLI) for Cameroon Pidgin English and Camfranglais. The dataset includes 228 spoken NLI items, recorded by a native speaker, with premise audio, text, NLI labels, and English glosses. Initial experiments using XLM-RoBERTa showed limited accuracy, with results not significantly better than chance and heavily influenced by hypothesis wording. Furthermore, speech recognition models like Whisper Large V3 struggled to accurately transcribe the Pidgin recordings, misclassifying them as other languages. AI
IMPACT Highlights challenges in developing NLP resources for under-resourced languages and the need for robust speech recognition models.
RANK_REASON The item describes a new academic dataset and associated research findings for natural language processing. [lever_c_demoted from research: ic=1 ai=1.0]
- Cameroon
- Cameroon Pidgin
- Camfranglais
- CamSpokenNLI
- Hugging Face
- Swahili
- textual entailment
- Whisper
- Whisper Large V3
- XLM-RoBERTa
- Yoruba
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →