p988744/eland-entity-sentiment-zh

Eland Entity Sentiment Chinese Dataset

A Chinese entity-level sentiment analysis dataset for training models to classify sentiment toward specific entities (brands, products, companies, people) mentioned in text.

Dataset Description

This dataset contains Chinese text samples with entity-level sentiment annotations. Unlike overall sentiment analysis, entity sentiment focuses on the sentiment expressed toward a specific entity mentioned in the text.

Key Features

• Entity-Specific Labels - Each sample contains a target entity and its sentiment
• Multi-Entity Scenarios - Same text can have different sentiments for different entities
• Financial Domain - Primarily focused on financial news and forum discussions
• Traditional Chinese - Taiwan-style Traditional Chinese text

Example

{
  "text": "台積電營收創新高，但聯電表現不如預期",
  "entity": "台積電",
  "entity_sentiment": "正面"
}

The same text with different entity:

{
  "text": "台積電營收創新高，但聯電表現不如預期",
  "entity": "聯電",
  "entity_sentiment": "負面"
}

Dataset Structure

Files

File	Samples	Description
train.jsonl	333	Training samples
test.jsonl	100	Test samples
benchmark/all_samples.jsonl	100	Benchmark evaluation set
chat_format/train.jsonl	333	Chat format for training
chat_format/test.jsonl	100	Chat format for evaluation

Data Fields

Field	Type	Description
text	string	The input text containing entity mentions
entity	string	The target entity to analyze sentiment for
entity_sentiment	string	Sentiment label: 正面, 負面, or 中立
category	string	Sample category (optional): standard, multi_entity, implicit, sarcasm
source	string	Data source (optional)

Label Distribution

Label	Train	Test	Total
正面 (Positive)	~111	~33	~144
負面 (Negative)	~111	~33	~144
中立 (Neutral)	~111	~34	~145
Total	333	100	433

Usage

Loading with HuggingFace Datasets

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("p988744/eland-entity-sentiment-zh")

# Access splits
train = dataset['train']
test = dataset['test']

# Example iteration
for sample in train:
    print(f"Text: {sample['text']}")
    print(f"Entity: {sample['entity']}")
    print(f"Sentiment: {sample['entity_sentiment']}")
    break

Loading JSONL Directly

import json

samples = []
with open('train.jsonl', 'r', encoding='utf-8') as f:
    for line in f:
        samples.append(json.loads(line))

Chat Format for Training

The chat_format/ directory contains pre-formatted data for instruction-tuned models:

{
  "messages": [
    {
      "role": "system",
      "content": "你是一個專業的實體情感分析助手..."
    },
    {
      "role": "user",
      "content": "分析以下文本對「台積電」的情感傾向。\n\n文本：台積電營收創新高\n\n請以 JSON 格式回答..."
    },
    {
      "role": "assistant",
      "content": "{\"entity_sentiment\": \"正面\"}"
    }
  ]
}

Benchmark Categories

The benchmark set (benchmark/all_samples.jsonl) includes challenging samples:

Category	Count	Description
standard	~50	Standard entity sentiment samples
multi_entity	~25	Multiple entities with different sentiments
implicit	~15	Implicit/indirect sentiment expression
sarcasm	~10	Sarcastic or ironic expressions

Model Trained on This Dataset

Model	Accuracy	Macro-F1	Repository
Eland Entity Sentiment	72.00%	70.80%	p988744/eland-entity-sentiment-zh

Model Variants

Version	Repository	Use Case
LoRA Adapter	p988744/eland-entity-sentiment-zh	HuggingFace + PEFT
GGUF	p988744/eland-entity-sentiment-zh-gguf	Ollama / llama.cpp
Full Merged	p988744/eland-entity-sentiment-zh-vllm	vLLM

Related Datasets

Dataset	Task	Repository
Eland Sentiment Data	Overall/Opinion Sentiment	p988744/eland-sentiment-zh-data
Eland Stance Data	Stance Detection	p988744/eland-stance-zh-data

Token Statistics

Split	Samples	Estimated Tokens
train	333	~59K
test	100	~18K
Total	433	~77K

Token counts estimated using Qwen3-4B tokenizer.

Calculate Tokens

from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("p988744/eland-entity-sentiment-zh")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

def count_tokens(split):
    total = 0
    for sample in dataset[split]:
        # Count text + entity
        total += len(tokenizer.encode(sample["text"], add_special_tokens=False))
        total += len(tokenizer.encode(sample["entity"], add_special_tokens=False))
    return total

print(f"Train tokens: {count_tokens('train'):,}")
print(f"Test tokens: {count_tokens('test'):,}")

Limitations

• Domain: Primarily financial domain, may not generalize well to other domains
• Size: Limited training data (333 samples)
• Language: Traditional Chinese only

Citation

@dataset{eland-entity-sentiment-zh,
  author = {Eland AI},
  title = {Eland Entity Sentiment Chinese Dataset},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/p988744/eland-entity-sentiment-zh}
}

License

Apache 2.0