alvations/usaar-wikicorpus-2015

USAAR Wikicorpus (2015)

Domain-specific Wikipedia sentence collections released by the Universität des Saarlandes (USAAR) for the SemEval hyponym/hypernym tasks. For each terminology entry (a single word or multiword expression) the corpus provides the set of Wikipedia sentences that mention it, grouped together as one document.

This release stitches the original sentence-blocks into one coherent paragraph per term (light deterministic cleanup, no LLM rewriting) and packages the four domains as a multi-config HF dataset.

Configs

Config	Docs	Raw sentences	Cleaned sentences	Avg chars/doc	Parquet
chemical	17,584	601,981	587,460	~4,890	47.6 MB
equipment	612	112,928	111,692	~26,560	9.5 MB
food	1,555	572,451	559,811	~49,180	45.8 MB
science	452	293,948	285,915	~97,040	23.1 MB
all	20,203	1,581,308	1,544,878	—	—

all is a glob over the four domain shards.

Usage

from datasets import load_dataset

ds = load_dataset("alvations/usaar-wikicorpus-2015", "chemical", split="train")
print(ds[0]["term"])             # 'agarose'
print(ds[0]["text"][:200])
# 'Agarose 3-glycanohydrolase may refer to one of two enzymes: Alpha-agarase (...

# all four domains in one go
all_ds = load_dataset("alvations/usaar-wikicorpus-2015", "all", split="train")
print(len(all_ds), "docs total")

Schema

Column	Type	Description
doc_id	int32	0-indexed document id, dense within each domain (matches the integer prefix in the original TSV header lines).
term	string	The terminology entry — single word or multiword expression (always lowercased in source).
text	string	Cleaned paragraph stitching the Wikipedia sentences that mention the term.
num_sentences_raw	int32	Number of non-empty source lines in the original block (before cleanup).
num_sentences_clean	int32	Number of sentences in the final paragraph (after cleanup).
domain	string	"chemical" / "equipment" / "food" / "science".

Cleanup pipeline

Each per-term block (one or more raw sentences) is normalised by a small deterministic pipeline before being joined into a paragraph. No LLM rewriting is applied — the original wording is preserved verbatim.

1. Strip + drop empty. Per-line whitespace stripped; empty lines dropped.
2. Drop noise lines. Pure URL lines (http://… or [http…) and lines shorter than 5 characters are dropped.
3. Dedupe consecutive duplicates. If two adjacent lines are byte-identical, the second is dropped.
4. Strip bare-title prefixes. A short line with no terminal punctuation that is a strict prefix of the next line (e.g. Agarose 3-glycanohydrolase followed by Agarose 3-glycanohydrolase may refer to ...) is dropped — these are Wikipedia section titles immediately followed by their full opening sentence.
5. Punctuation + join. Each remaining line is given a trailing . if it doesn't already end in ., !, ?, :, ;, ", ’, ”, or ). The cleaned sentences are then joined with a single space to form one paragraph per term.

Net cleanup is ~1–3% of raw lines per domain. Both raw and clean counts are kept as columns so you can audit the loss.

Source format (for reference)

The upstream release was four plain-text files (WIKI_chemical.txt, WIKI_equipment.txt, WIKI_food.txt, WIKI_science.txt) shaped as repeated blocks of:

0<TAB>agarose
Agarose 3-glycanohydrolase
Agarose 3-glycanohydrolase may refer to one of two enzymes:
Alpha-agarase (...) is an enzyme with system name "agarose 3-glycanohydrolase".
1<TAB>nickel sulfate heptahydrate
Nickel sulfate heptahydrate is ...
...

Citation

This corpus was assembled by the USAAR team to support their SemEval hyponym/hypernym systems. If you use it, please cite both papers:

@inproceedings{tan-etal-2016-usaar,
    title     = "{USAAR} at {S}em{E}val-2016 Task 13: Hyponym Endocentricity",
    author    = "Tan, Liling  and Bond, Francis  and van Genabith, Josef",
    booktitle = "Proceedings of the 10th International Workshop on Semantic Evaluation ({S}em{E}val-2016)",
    month     = jun,
    year      = "2016",
    address   = "San Diego, California",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/S16-1203/",
    doi       = "10.18653/v1/S16-1203",
    pages     = "1303--1309",
}

@inproceedings{tan-etal-2015-usaar-wlv,
    title     = "{USAAR}-{WLV}: Hypernym Generation with Deep Neural Nets",
    author    = "Tan, Liling  and Gupta, Rohit  and van Genabith, Josef",
    booktitle = "Proceedings of the 9th International Workshop on Semantic Evaluation ({S}em{E}val 2015)",
    month     = jun,
    year      = "2015",
    address   = "Denver, Colorado",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/S15-2155/",
    doi       = "10.18653/v1/S15-2155",
    pages     = "932--937",
}

License

Source content is derived from English Wikipedia and inherits its license: CC BY-SA 3.0. Reuse must preserve attribution to the Wikipedia contributors and the ShareAlike requirement.

Provenance of this release

• Raw upstream files (timestamped 2015-04-14): WIKI_chemical.txt, WIKI_equipment.txt, WIKI_food.txt, WIKI_science.txt.
• Conversion: deterministic cleanup pipeline described above; sentences joined into one paragraph per (domain, doc_id); one parquet shard per domain.
• No row filtering, semantic deduplication, or external knowledge injection applied.