bjorndev/efik-text-corpus

Efik Text Corpus

A collection of 2,827 unique Efik text records, including 201 entries with source-provided English counterparts, published by bjorndev so others can inspect, correct, and build on the collected material.

The collection contains prose, hymn versions, prayers, dictionary terms, numbers, and conversational phrases. It preserves source spellings and Unicode diacritics, with per-entry provenance and quality flags. English text was extracted from the supplied sources; no translations were generated.

Quick start

from datasets import load_dataset

corpus = load_dataset("bjorndev/efik-text-corpus", split="train")
pairs = load_dataset("bjorndev/efik-text-corpus", "english_pairs", split="train")

The configurations use CSV files, where sources and quality_flags contain JSON-encoded arrays. Parse them with json.loads when needed. To preserve empty strings instead of CSV nulls, pass keep_default_na=False to load_dataset. The downloadable JSONL files retain native arrays and heterogeneous raw source values; they can be read with Python's standard json module.

Both configurations have a single train split for distribution convenience. No independent training/evaluation split or benchmark validation is implied. The English-paired configuration is a subset of the complete corpus.

Source attribution and rights

Source collections are identified by their supplied archive/file names and internal paths in each record's sources field and source_inventory.jsonl. Original upstream URLs, authorship, and licensing were not established for every source. No blanket license is assigned to the underlying texts. Rights remain with their respective holders; users should verify the relevant permissions for their intended use. Publication does not certify that all texts are freely licensed.

Limitations

This is an extracted, unverified collection. It is heavily weighted toward religious material and is not representative of all everyday Efik usage. OCR noise, source spelling errors, inconsistent orthographies, and uncertain example alignments remain. Some app content and small UI/test strings are included with flags. English alternate hymn titles are metadata, not translations of the full lyrics. Related-language content may require review by Efik speakers. No Reddit content is included.

The JSONL/CSV data files are exact copies of the local extracted dataset. The extraction script is included for inspection; rebuilding requires the original source files, which are not bundled in this release.

Detailed dataset documentation

Extracted from the files supplied in this folder. No translations were generated and no external content was downloaded. Original input files were left untouched.

Files

• efik_dataset.csv: complete dataset; opens in Excel and other spreadsheet tools.
• efik_dataset.jsonl: the same dataset, one JSON object per line, with structured provenance.
• efik_english_pairs.csv / efik_english_pairs.jsonl: entries with English text supplied by the sources.
• source_inventory.jsonl: every scanned file, including archive members, its hash, size, and extraction disposition.
• excluded_dictionary_rows.jsonl: four dictionary page headers excluded from the dataset, with their original values.
• summary.json: counts by content type and quality flag.

There are 2,827 unique records, approximately 140,987 whitespace-separated Efik-field words, and 201 records with English. Word counts include titles, stanza numbers, source spelling variants, and occasional source annotations. These are extraction counts, not counts of independently verified translations or distinct hymn compositions.

Content	Unique records
Prose paragraphs	1,196
Hymn texts/versions	807
Hymn titles/index entries	494
Dictionary terms	132
Dictionary examples	29
Numbers/vocabulary	20
Conversational phrases	21
Prayer passages	79
Prayer headings	29
Scripture epigraphs	16
Efik site names/test text	4

Columns

• id: deterministic content-derived identifier.
• efik_text: extracted text, retaining source spellings and diacritics.
• english_text: source English counterpart; empty when none is available.
• translation_status: source_provided or not_available. This describes availability, not linguistic correctness.
• type: content category from the table above.
• title: source Efik hymn title where available.
• english_title: an English alternate hymn title from application metadata. This is not a full translation of the hymn lyrics, and is deliberately separate from english_text.
• notes: dictionary usage notes or explanations of extraction decisions. Some notes contain further mixed-language examples retained verbatim.
• quality_flags: issues/context requiring review.
• sources: all source occurrences, with file path and a resource name, block, row, or line locator. Original dictionary cells, phrase cells, and prose paragraphs are also retained here.

sources and quality_flags are JSON arrays inside CSV cells. JSONL preserves their native structure. Archive paths use archive.zip::path/inside/archive. Text-file line locators count newlines after converting CR and CRLF to LF.

Extraction and deduplication

All 11 ZIP archives, both top-level text files, and the already-unpacked prayer folder were inspected: 1,119 files/archive members, with corpus text extracted from 450. The duplicate hymn-book ZIP and repeated prayer folders contribute provenance without duplicating identical dataset records.

Records are deduplicated by the exact Efik/English text pair after NFC Unicode normalization and whitespace cleanup. Different spellings, punctuation, case, and hymn versions are preserved. Prose is segmented by blank lines, with wrapped lines joined. Its 1,205 source paragraphs become 1,196 unique records, with every occurrence retained. Hymns retain their stanza markers and line order; source choruses are included. Hymn titles can occur as separate records and as metadata.

English hymn titles are retained only as metadata; no alignment of English titles to full Efik lyrics is asserted. Missing English remains empty. Explicit source English–Efik vocabulary and simple example pairs are included. Obvious misplaced CSV columns and greeting labels were repaired with original cells retained and flags added. OCR spelling errors and merged words were not silently rewritten. The orphan dictionary term with no English headword is retained without English.

Quality and coverage limits

The paired subset needs language review before use as a translation benchmark. In particular, dictionary examples contain OCR spacing/diacritic errors, and some English headwords are truncated in the source.

Useful flags include:

• source_ocr_noise: prose/dictionary sources may contain OCR errors.
• recovered_shifted_columns: an explicit term or phrase was moved out of the wrong source field.
• truncated_english_headword / missing_english_headword: damaged or absent English in the source.
• example_alignment_needs_review: source usage-example alignment needs review.
• unverified_app_content: hymn data from the newer app, whose linguistic quality and provenance are not established by the supplied files.
• index_or_commented_source: hymn incipits recovered from app index lists, including commented-out entries.
• ui_or_test_fixture: small Efik site-name/test strings, separate from prose and hymns.
• mixed_language_reference: a source passage includes a Bible citation or English fragment.
• english_annotations_moved_to_notes: greeting/reply labels removed from the Efik field and preserved in notes.

The nwed-nyin-web-master.zip website retrieves dictionary/articles from a remote service; that database is not included in the archive. Only its locally present Efik strings could be extracted. The archive also contains test words in other languages; these were not treated as Efik.

Images, icons, fonts, audio, compiled files, and project infrastructure are listed in the inventory but were not OCRed or transcribed. The prayer illustrations are accompanied by machine-readable HTML passages. This dataset covers locally recoverable text; it does not claim transcription of every binary asset.

Source rights are not changed by extraction; this compilation does not assign a new license to the underlying texts.

Rebuild and use

From the parent folder, run:

python3 scripts/build_dataset.py

The extractor uses Python's standard library, reads ZIP members directly, and does not execute application code from the archives. Rebuilding overwrites the generated CSV, JSONL, and summary files in dataset/.

import json

with open('dataset/efik_dataset.jsonl', encoding='utf-8') as f:
    records = [json.loads(line) for line in f]

pairs = [r for r in records if r['english_text']]
prose = [r for r in records if r['type'] == 'prose']

Validation checked unique IDs/pairs, nonempty text and provenance, CSV/JSONL agreement, extraction of all 366 JSX hymn files, preservation of all non-whitespace characters in the prose source after NFC normalization, and accounting for every dictionary row.