p988744/eland-ner-zh

Eland NER Dataset - Chinese Public Opinion NER

Chinese Named Entity Recognition dataset for public opinion analysis, containing 1,979 annotated articles with 9 entity types. Available in two configurations: raw annotations and BERT-tokenized format.

Dataset Description

This dataset is designed for training NER models to extract entities from Chinese news articles, social media posts, and forum discussions, particularly in the context of public opinion monitoring in Taiwan.

Entity Types

Type	Label	Description	Examples
PER	人物	Named persons	川普, 魏哲家, 賴清德
ORG	組織機構	Organizations	台積電, 民進黨, FBI
LOC	地點	Locations	台北, 美國, 信義區
DATE	時間	Temporal expressions	2026年, Q4, 春節
EVENT	事件	Named events	CES 2025, 股東會
PROD	產品/品牌/服務	Products/brands/services	iPhone, ChatGPT
LAW	法規	Laws and regulations	證券交易法, 勞基法
METRIC	指標名稱	Metric names	營收, 本益比, GDP
TOPIC	輿情主題詞	Topic keywords	AI, 半導體, 碳中和

Dataset Configurations

Config: raw (Original Annotations)

Human-readable format with article text and entity spans.

Fields:

Field	Type	Description
id	string	Unique article identifier
title	string	Article title
content	string	Article body text
entities	string (JSON)	Entity annotations: [{"text": "...", "type": "..."}]
source	string	Data source (news or bbs)

Splits:

Split	Samples	Description
train	1,880	95% random split (seed=42)
validation	99	5% random split (seed=42)

Config: bert (BERT Tokenized)

Pre-tokenized format ready for BertForTokenClassification training.

Fields:

Field	Type	Description
input_ids	list[int]	Token IDs from hfl/chinese-roberta-wwm-ext
attention_mask	list[int]	Attention mask (1=real, 0=padding)
labels	list[int]	BIO label IDs (-100=ignore, 0=O, 1-18=B/I entity labels)

Splits:

Split	Samples	Description
train	1,619	From 95% annotated data
validation	86	From 5% annotated data
test	49	Separate manual test set

BIO Label Mapping:

Label	ID	Label	ID
O	0	B-PROD	11
B-PER	1	I-PROD	12
I-PER	2	B-LAW	13
B-ORG	3	I-LAW	14
I-ORG	4	B-METRIC	15
B-LOC	5	I-METRIC	16
I-LOC	6	B-TOPIC	17
B-DATE	7	I-TOPIC	18
I-DATE	8
B-EVENT	9
I-EVENT	10

Entity Distribution (Full Dataset)

Type	Count	Percentage
ORG	3,519	26%
PROD	3,202	24%
TOPIC	2,690	20%
PER	1,849	14%
LOC	1,809	14%
DATE	1,547	11%
EVENT	749	6%
METRIC	556	4%
LAW	190	1%
Total	~13,100

Average: ~6.6 entities per article.

Example Samples

Raw Format

{
  "id": "news_1770767960947_N01",
  "title": "開春美國票房黑馬《鐵肺》 300萬美金成本收入近4000萬",
  "content": "改編自同名知名恐怖遊戲的美國獨立製片《鐵肺》...",
  "entities": [
    {"text": "鐵肺", "type": "PROD"},
    {"text": "美國", "type": "LOC"},
    {"text": "Markiplier", "type": "PER"},
    {"text": "票房", "type": "METRIC"},
    {"text": "恐怖遊戲", "type": "TOPIC"}
  ],
  "source": "news"
}

BERT Format

{
  "input_ids": [101, 7030, 4989, 5765, ..., 102],
  "attention_mask": [1, 1, 1, 1, ..., 1],
  "labels": [-100, 3, 4, 4, 0, 0, 1, 2, ..., -100]
}

Usage

Loading the Dataset

from datasets import load_dataset

# Raw annotations
ds_raw = load_dataset("p988744/eland-ner-zh", "raw")
print(ds_raw["train"][0])

# BERT tokenized format
ds_bert = load_dataset("p988744/eland-ner-zh", "bert")
print(ds_bert["train"][0])

Training with BERT Format

from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForTokenClassification,
    TrainingArguments,
    Trainer,
)

dataset = load_dataset("p988744/eland-ner-zh", "bert")
model = AutoModelForTokenClassification.from_pretrained(
    "hfl/chinese-roberta-wwm-ext", num_labels=19
)

# Dataset is already tokenized - use directly with Trainer
trainer = Trainer(
    model=model,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    # ... your training arguments
)

Source

• News articles: Taiwan financial and general news
• BBS/Forum posts: PTT, Dcard, and other forums
• Annotation: LLM-assisted with human review

Related Resources

Models

Model	Repository
Eland NER	p988744/eland-ner-zh

Related Datasets

Dataset	Task	Repository
Sentiment	Sentiment Analysis	p988744/eland-sentiment-zh
Stance	Stance Detection	p988744/eland-stance-zh
Entity Sentiment	Entity Sentiment	p988744/eland-entity-sentiment-zh

Citation

@misc{eland-ner-zh-data,
  author = {Eland AI},
  title = {Eland NER Dataset: Chinese Public Opinion Named Entity Recognition},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/p988744/eland-ner-zh}
}

License

Apache 2.0