VelkroLM/hausa-stopwords-corpus

Hausa Stopword Candidates and Frequency Scores

A reproducible Hausa lexical resource containing frequency-scored stopword candidates. This repository is organized for inspection, preprocessing experiments, and future Hausa-speaker review. It does not publish a final stopword list or a final human-reviewed stopword count.

Quick navigation

Need	Go to
Browse candidates in the Dataset Viewer	data/hausa_stopword_candidates.jsonl
Efficient analysis	derived/hausa_stopword_candidates.parquet
Read the scoring method	docs/METHODOLOGY.md
Understand the cleaned schema	docs/SCHEMA.md
Follow the human-review process	docs/HUMAN_REVIEW.md
Read the cleanup report	docs/CLEANUP_REPORT.md
Inspect machine-readable metrics	metrics/
Inspect the original scored inventory	raw_scores/full_word_scores.json
Browse ranked compact batches	batches/
View the top 1,000 candidates	top1000/
Review the unpublished-count placeholder	final_stopwords/FINAL_STOPWORDS_COUNT.txt

Repository tree

.
├── README.md                         # This overview and navigation page
├── .gitattributes                    # Repository/LFS settings
├── data/
│   └── hausa_stopword_candidates.jsonl  # Dataset Viewer source; one schema per line
├── derived/
│   └── hausa_stopword_candidates.parquet # Columnar copy for analysis
├── batches/
│   ├── batch_0001.json ... batch_0109.json # Ranked compact candidate batches
├── raw_scores/
│   └── full_word_scores.json          # Original complete frequency inventory
├── top1000/
│   ├── top_1000_candidates.csv
│   ├── top_1000_candidates.json
│   └── top_1000_candidates.txt
├── metrics/
│   ├── normalization_metrics.json     # Counts and review-flag metrics
│   └── cleanup_validation.json        # Structural and viewer-style validation
├── docs/
│   ├── INDEX.md                       # Documentation map
│   ├── METHODOLOGY.md                 # Original scoring method
│   ├── SCHEMA.md                      # Normalized record schema
│   ├── HUMAN_REVIEW.md                # Review workflow; final count remains pending
│   └── CLEANUP_REPORT.md              # Normalization and viewer-repair report
└── final_stopwords/
    └── FINAL_STOPWORDS_COUNT.txt      # Deliberate placeholder; no final count yet

Status and scope

This release is a cleaned candidate inventory. All records remain marked human_review_pending. The final stopword list and final reviewed count will be created only after qualified Hausa-speaker review.

The source methodology pools four named source families: CC-100 Hausa web text, Hausa Wikipedia, Leipzig Corpora Collection Hausa Community 2017, and QA/instruction pairs from rufatronics/mt5_hausa. Scores are pooled across the sources; the current composite score equals the document-frequency ratio.

Current metrics

Metric	Value
Candidate records	325,537
Total term frequency	16,911,109
Maximum document frequency	316,137
Composite-score range	0.0–0.56562
English-marker review flags	30
Name review flags	34
News-bias review flags	31
Records with any review flag	92
Malformed source entries	0
Final reviewed stopword count	Not published

The flags are review aids, not automatic deletion rules. For example, common Hausa grammatical words such as da, na, ta, and ya may be frequent and linguistically important. Frequency alone does not make a word a stopword.

Intended use

Use this repository for tokenizer and normalization experiments, lexical analysis, frequency-aware sampling, candidate-list comparison, and controlled preprocessing research. For generative Hausa training, keep the candidate list separate from the text corpus and do not remove frequent words without linguistic review.

Quality and limitations

The normalized JSONL uses one consistent 13-field schema and is explicitly assigned to the Dataset Viewer train split. The Parquet copy has the same 325,537 rows. Local loader tests and the live Dataset Viewer rows endpoint both returned valid records.

The resource remains limited by pooled source bias, incomplete source-level weighting, heuristic English/name/news flags, uneven dialect coverage, and the absence of final Hausa-speaker adjudication. The candidate count is therefore not a final stopword count.

License and provenance

The repository metadata declares MIT for the published derived artifacts. The underlying source corpora may have separate terms. Review each source license before redistributing raw source material or using derived artifacts commercially.

For the complete technical details, see docs/INDEX.md.