Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
567k
# Bosnian Corpus (v1.0)
This package contains the plain-text **Bosnian Corpus (v1.0)** and the
accompanying documentation papers.
- **DOI:** https://doi.org/10.5281/zenodo.17757098
- **Author:** Hasan Kahrimanović
- **Release date:** 2025-11-29
The corpus is built specifically for quantitative studies of Bosnian
(language entropy, “language energy”) and for general NLP experiments.
---
## 1. Contents of this package
The main files in this archive are:
- `bosnian_corpus.txt`
Plain-text Bosnian corpus (approx. 6.03 GB).
One sentence per line, cleaned and deduplicated.
- `bosnian_corpus_en.pdf`
English description of the corpus construction and preprocessing pipeline.
- `bosanski_korpus_bs.pdf`
Bosnian-language description of the same corpus.
Optional (only if included):
- `bosnian_corpus_*.txt`
Genre-specific subsets (NEWS, OPINION, FORUM_CHAT, INFO_HOWTO,
LEGAL_ADMIN, LITERATURE, ADS_PROMO, MIX_OTHER).
---
## 2. Data format
- **Encoding:** UTF-8, Unicode NFC.
- **Line structure:** one sentence or short message per line.
- **Tokenisation:** white-space separated, punctuation attached to tokens
as in standard text.
- **Script:** Latin alphabet; Cyrillic and heavy non-Bosnian segments are
strongly reduced during cleaning.
- **Size:**
- bytes: ~6,182,905,888
- tokens: 942,515,845
- types (vocabulary): 4,463,058
For detailed statistics and cleaning steps, see the PDFs in this package.
---
## 3. Provenance of the text
The corpus is derived from three publicly available resources hosted on
the CLARIN.SI repository:
1. **Sarajevo Corpus of SMS Messages in Bosnian 1.1**
2. **Bosnian web corpus bsWaC 1.1**
3. **Bosnian web corpus CLASSLA-web.bs 1.0**
All source corpora are cited in the accompanying papers and should be
cited separately by users of this dataset.
---
## 4. Cleaning and processing (short summary)
The following high-level steps were applied:
- Conversion of XML/vert/CONLLU formats to plain text (one sentence per line).
- Unicode normalisation (NFC), UTF-8 output.
- Removal of technical noise (URLs, boilerplate, CMS headers/footers,
file names, log fragments, etc.).
- Heuristic removal of lines with a high proportion of non-letter symbols.
- Language filtering to favour Bosnian (diacritics, function words).
- Genre classification into several super-genres using a TF–IDF +
logistic regression model trained on CLASSLA-web.bs genres.
- Additional *paper-grade* cleaning and optional deduplication.
Full details and references are provided in `bosnian_corpus_en.pdf`.
The Python pipeline used for these steps is available at:
> https://github.com/H4sK0/bosnian-corpus-pipeline
---
## 5. License
The **text corpus** in this package is released under:
> **Creative Commons Attribution–ShareAlike 4.0 International (CC BY-SA 4.0)**
You are free to:
- share — copy and redistribute the material in any medium or format
- adapt — remix, transform, and build upon the material
End of preview. Expand in Data Studio

Bosnian Corpus v1.0 (cleaned)

This dataset provides a cleaned and genre-annotated corpus of contemporary Bosnian, designed for quantitative analysis of language entropy, “language energy”, and modern NLP tasks.

The canonical release of this corpus is published on Zenodo: DOI: 10.5281/zenodo.17757098


Corpus composition

The corpus is built from three publicly available resources released via the CLARIN.SI repository:

  1. Sarajevo Corpus of SMS Messages in Bosnian 1.1
  2. Bosnian Web Corpus bsWaC 1.1
  3. Bosnian Web Corpus CLASSLA-web.bs 1.0

All sources were converted to plain text, cleaned, normalized, partially deduplicated, and merged into a single consistent dataset.


Size and statistics

  • Total size: ~6.18 GB (≈ 6,182,905,888 bytes)
  • Lines: ~46,258,935
  • Tokens: ~942,515,845

Genre structure

The web portion of the corpus is organized into the following “super-genres”:

  • News
  • Opinion
  • Forum / Chat
  • Info / HowTo
  • Legal / Administrative
  • Literature
  • Ads / Promo
  • Mix / Other

For each super-genre, a separate text file is provided, along with one global file concatenating all genres. The global file is intended for entropy estimation and language-model training.


Cleaning and normalization

Cleaning focuses on removing technical noise that would bias frequency distributions and entropy estimates, while preserving the linguistic signal:

  • Unicode normalization (UTF-8, NFC)
  • Correction of common mojibake artifacts
  • Removal of URLs, e-mail addresses, file names, boilerplate, CMS/navigation lines
  • Filtering of lines with a high proportion of non-letter characters
  • Optional digit normalization and lowercasing
  • Language filtering to keep primarily Bosnian text

Files

  • bosnian_corpus_all.txt — full corpus (all genres combined)
  • Per-genre text files:
    • news
    • opinion
    • forum_chat
    • info_howto
    • legal_admin
    • literature
    • ads_promo
    • mix_other
  • README.txt — dataset description

Associated research papers (Bosnian and English) are published separately on Zenodo.

  • Download: data/bosnian-corpus-1.0.zip

Code availability

Preprocessing, cleaning, and entropy-calculation scripts are available on GitHub:

https://github.com/H4sK0/bosnian-corpus-pipeline


License

This dataset is released under the
Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

Users must:

  • credit this Zenodo record and the original source corpora
  • distribute derivative datasets under the same or a compatible license

Citation

@dataset{kahrimanovic2025bosniancorpus,
  title   = {Bosnian Corpus (v1.0): Cleaned Web and SMS Text for Entropy and NLP Research},
  author  = {Kahrimanović, Hasan},
  year    = {2025},
  publisher = {Zenodo},
  doi     = {10.5281/zenodo.17757098}
}
Downloads last month
26