IgboNLP-Research/IgboAI

Open research on AI for Igbo language: self-updating literature tracking, corpora, and benchmarks spanning NLP, speech, and language models for a low-resource, tonal language. Tracks new papers and models nightly, building corpora, and benchmarking across NLP, ASR/TTS, and LLMs. # IgboAI Open research on AI for the Igbo language: self-updating literature tracking, corpora, and benchmarks spanning NLP, speech, and language models for a low-resource, tonal language. > **Status:** early infrastructure stage. Literature tracking is live; corpora and benchmarks are planned. Nothing here should be treated as a stable resource or citable result yet. > > **Licensing:** Licensing differs by artefact: code Apache-2.0, data per-source, names and logos reserved. See Licensing. ## What lives here - `RELATED_WORK.md` - a running, dated log of new papers, models, and datasets relevant to Igbo and closely transferable African/low-resource language work. Updated by automation, curated by a human at the PR stage. - `scripts/fetch_candidates.py` - the deterministic fetch layer (see below). - `.github/workflows/literature-tracking.yml` - the nightly automation. - Corpora and benchmark suites will be added as the project grows. ## How the automation works Every night, a scheduled GitHub Actions workflow keeps the literature log up to date. It is deliberately split into two layers: **Deterministic fetch, LLM curation.** A stdlib-only Python script queries three sources - the arXiv API (preprints), OpenAlex (venue-published work: ACL Anthology venues such as ACL, EMNLP, EACL, COLING, LREC and TACL, plus workshops and journals), and the Hugging Face Hub (models and datasets). It filters by date, deduplicates against a history file (`.github/tracking/seen.json`), and writes the candidates to a scratch JSON file. Only then, and only if there is anything new, does Claude Code run: it judges relevance to Igbo specifically, writes researcher-oriented summaries into `RELATED_WORK.md`, and opens a pull request. Retrieval failures and summarization failures are therefore easy to distinguish, and days with nothing new cost nothing. **Human-in-the-loop at the merge boundary.** The automation never pushes to `main`. Everything arrives as a pull request for human rev …