Geonwoohong/aihub-webcorpus-morph-train-tokenized-ko

Dataset Description

This dataset contains the AIHub Korean Web Corpus cleaned and morphologically annotated with the word identifier.
Each record stores morpheme-level tokens and two subsets: semantic and stylistic.

---

Dataset Structure

Data Fields

Field	Type	Description
text	string	Original sentence reconstructed from morphemes
semantic	list[dict]	Subset of content-bearing morphemes
stylistic	list[dict]	Subset of grammatical/stylistic morphemes

Example

{
  "text": "정부가 LTE 서비스 도입 등...",
  "semantic": [
    {"form": "정부", "tag": "NNG"},
    {"form": "LTE", "tag": "SL"},
    {"form": "도입", "tag": "NNG"},
    {"form": "추진", "tag": "NNG"}
  ],
  "stylistic": [
    {"form": "가", "tag": "JKS"},
    {"form": "을", "tag": "JKO"},
    {"form": "한다", "tag": "VX+EF"}
  ],
}

Tag Sets

• Semantic (examples):
  NNG, NNP, NNB, NR, NP, VV, VA, VX, VCP, VCN, XR, MM, MAG, MAJ, SN, SL, XPN, and USER*

• Stylistic (examples):
  JKS, JKC, JKG, JKO, JKB, JKV, JKQ, JX, JC, EP, EF, EC, ETN, ETM, IC, XSN, XSV, XSA, XSM

• Dropped (examples):
  SF, SP, SS, SSO, SSC, SE, SO, SW, plus W_URL, W_EMAIL, W_HASHTAG, W_MENTION, W_SERIAL, W_EMOJI

---

Data Splits

• Split: train
• Composition: Consists of user-prepared Apache Arrow shards (e.g., data-00000-of-XXXXX.arrow),
  with additional Parquet indexing and cache structures automatically managed by 🤗 Datasets after upload.

---

Source Data

• Original dataset: AIHub Korean Web Corpus (2021)
• Tokenizer / POS tagger: Kiwi (Kiwipiepy, CoNg model)

---

Preprocessing

• The original AIHub corpus was cleaned by removing HTML tags, URLs, emails, mentions, and hashtags,
  normalizing punctuation and spacing, and filtering out overly short or long samples.
• Each document was morphologically analyzed using Kiwi (CoNg), producing morpheme-level annotations
  separated into semantic (content-bearing) and stylistic (grammatical) subsets.
• The processed data was saved as Apache Arrow shards for upload to the Hub.

---

Usage

from datasets import load_dataset

# Replace with your repo id, e.g., "Geonwoohong/ko-web-kiwi-morph"
ds = load_dataset("YOUR_USERNAME/YOUR_REPO", split="train")

print(ds[0]["text"])
print(ds[0]["semantic"][:5])
print(ds[0]["stylistic"][:5])

---

Data Files

• Stored as user-prepared Apache Arrow shards on the Hub.
• 🤗 Datasets automatically manages Parquet-based indexing and caching structures for efficient streaming and loading.
• You don’t need to reference individual files — simply load the dataset using its repository ID as shown in Usage.

---

Limitations

• Possible domain bias from the source AIHub web corpus.
• Morphological tagging quality depends on the Kiwi (CoNg) analyzer and may vary by domain or sentence style.
• Token-level boundaries can occasionally differ from other Korean analyzers (e.g., MeCab, Khaiii),
  so users should ensure consistency when combining datasets.

---

Citation

AIHub. Korean Web Corpus.
Available at: https://aihub.or.kr/aihubdata/data/view.do?dataSetSn=624

Kiwi. Korean Intelligent Word Identifier.
Available at: https://github.com/bab2min/Kiwi