airesearch/Bilingual_StockTBSA

Bilingual Target-Based Stock Sentiment Dataset (Thai-English)

Abstract / Motivation

This work introduces a new bilingual Target-Based Sentiment Analysis (TBSA) dataset focused on the stock market domain. We collected stock-related financial news from both Thai and international sources, totaling approximately 10,300 Thai and 10,120 English articles.

Each sentence is annotated at the target (TICKER) level using one of six sentiment labels: positive, negative, neutral, exclude, ambiguous, and not stock.

• The ambiguous class denotes unclear sentiment polarity but was excluded from modeling due to its rarity.

• The not stock class corresponds to ticker-like entities (e.g., indices or organizations) erroneously matched during the ticker extraction process. It is included in the dataset statistics for transparency but omitted from model training and evaluation.

Dataset Format

The released dataset is structured in wide format, with one row per news article.

Each article contains:

The full text of the article

A list of target stock mentions and their corresponding sentiment labels

{
  "Article_id": "1",
  "Data-source": "Prachachat",
  "Date": "2018-01-03",
  "Year": 2018
  "Text": "PACE ออกหุ้นเพิ่มทุน PP จำนวน 400 ล้านหุ้น ให้ SCB มูลค่ารวม 204 ลบ. ผู้สื่อข่าวรายงานว่า บมจ.เพซ ดีเวลลอปเมนท์ คอร์ปอเรชั่น (PACE) ...",
  "Ticker_sentiments": [
    {"ticker": "PACE", "sentiment": "positive"},
    {"ticker": "SCB", "sentiment": "positive"}
  ]
}

📌 Label schema: Although the released dataset includes six sentiment classes (positive, negative, neutral, exclude, ambiguous, not_stock), the experiments reported in our published paper use only four main classes: positive, negative, neutral, and exclude.

🗓️ Temporal splitting: In our published experiments, we adopt a temporal split for model training and evaluation as follows:

Train: 2018–2020

Validation: 2021

Test: 2022–2023

Users are free to perform their own data splitting as needed for different experimental setups.

Citation

If you use this dataset or code, please consider citing:
@misc{uthayopas2025tbsa,
  author={Uthayopas, Chayapat
          and Mai-On, Chalermpun
          and Phatthiyaphaibun, Wannaphong
          and Buaphet, Weerayut},
          and Sawatphol, Jitkapat},
          and Sae lim, Sitiporn},
          and Vongkulbhisal, Jayakorn},
          and Vorawathanabuncha, Jasarin},
          and Nutanong, Sarana},
          and Udomcharoenchaikit, Can},
  title = {Thai-English Target-Based Stock Sentiment Dataset for Financial News with ICL-Based Evaluation},
  year = {2025},
  howpublished = {\url{https://github.com/vistec-AI/Bilingual_StockTBSA}},
  note = {Preprint available soon}
}