Geonwoohong/modu-morph-train-encoded-ko

Dataset Description

This dataset is a morphologically encoded Korean corpus constructed for morph-related research.
It integrates eight cleaned Korean sources — written, spoken, nonpublication, newspaper, messenger, aspect_emotion, emotion, and dialogue —
and encodes each sample into structured morpheme sequences for semantic and stylistic streams.

Each sentence was analyzed using Kiwi (Korean Intelligent Word Identifier),
split into semantic, stylistic, and combined morpheme representations,
and encoded into numerical tensors stored in .npz shards for direct model loading.

---

Dataset Structure

Files

File Type	Example	Description
Vocabulary	form_vocab.json, tag_vocab.json, text_vocab.json	Morpheme, POS, and text token vocabularies (shared across shards, size = 64,000)
NPZ Shards	train_tensor.shard0001.npz – train_tensor.shard0023.npz	Packed arrays containing morpheme IDs and offsets for semantic/stylistic/text

Data Fields (inside each .npz)

Field	Type	Description
sem_form_values, sem_form_offsets	np.ndarray[int32, int64]	Semantic morpheme token IDs + offsets
sem_tag_values, sem_tag_offsets	np.ndarray[int32, int64]	Semantic POS tag IDs + offsets
sty_form_values, sty_form_offsets	np.ndarray[int32, int64]	Stylistic morpheme token IDs + offsets
sty_tag_values, sty_tag_offsets	np.ndarray[int32, int64]	Stylistic POS tag IDs + offsets
txt_form_values, txt_form_offsets	np.ndarray[int32, int64]	Combined morpheme token stream (for text reconstruction)
source_ids, source_strings	np.ndarray[int32], np.ndarray[str]	Source corpus metadata (domain index and name)
line_no	np.ndarray[int64]	Original line index within its source
uid64	np.ndarray[uint64]	64-bit hash for document identity (from blake2b)

---

Example Record

Each row corresponds to a single document entry.

{
  "uid64": 473217028191553817,
  "source": "AIHub Web Reviews",
  "semantic": [
    {"form": "제품", "tag": "NNG"},
    {"form": "이", "tag": "JKS"},
    {"form": "좋", "tag": "VA"},
    {"form": "아요", "tag": "EF"}
  ],
  "stylistic": [
    {"form": "이", "tag": "JKS"},
    {"form": "어요", "tag": "EF"}
  ]
}

---

Data Splits

• Split: train
• Size: ~5.8 M documents (≈ 23 NPZ shards, each 0.8–1.4 GB compressed)
• Composition: Merged from eight major public Korean sources.
  Each entry contains both semantic and stylistic morpheme streams.

---

Preprocessing

Morphological analysis and segmentation were performed via:

from kiwipiepy import Kiwi
kiwi = Kiwi(model_type="cong", num_workers=8, typos="basic_with_continual_and_lengthening")
tokens = kiwi.tokenize(text, normalize_coda=True)

Then split into semantic/stylistic streams:

SEMANTIC_TAGS = {"NNG","NNP","VV","VA","XR","MM","MAG","SN","SL"}
STYLE_TAGS    = {"JKS","JKO","JKB","EP","EF","EC","ETN","ETM","IC","XSN","XSV"}

Finally, the encoded dataset was generated via encode_morph.py:

python encode_morph.py \
  --jsonl_name train_morph.jsonl.gz \
  --min_freq_form 1 \
  --max_form_vocab 64000 \
  --shard_size 400000 \
  --tensor_out train_tensor.npz

---

Usage

import numpy as np, glob

# Load a shard
f = sorted(glob.glob("train_tensor.shard*.npz"))[0]
z = np.load(f, allow_pickle=False)

print(z["sem_form_values"].shape, z["sem_tag_values"].shape)
print(len(z["source_strings"]), "unique sources")

For Hugging Face integration:

from datasets import load_dataset

ds = load_dataset("Geonwoohong/modu-morph-encoded-ko", split="train", streaming=True)
for sample in ds.take(2):
    print(sample)

---

Notes

⚠️ The dataset viewer may be unavailable due to large shard sizes.
To inspect locally, use:

import numpy as np
z = np.load("train_tensor.shard0001.npz")
print(z.files)

• Vocabulary capped at 64,000 morpheme tokens.
• Morpheme-based structure optimized for dual-stream encoders (semantic + stylistic).
• Designed for representation learning, anomaly detection, and style disentanglement tasks.
• Encoded with Kiwi(CoNg) for robust Korean morphological decomposition.
• Each shard self-contained (no external metadata required).

---

Citation

---