gtfintechlab/Numclaim

Dataset Card for “NumClaim: Numerical Claim Detection in Finance”

---

Dataset Description

NumClaim is a sentence‑level corpus for detecting numerical claims in financial text.
A sentence is labelled as INCLAIM if it expresses a forward‑looking or speculative financial forecast, and OUTOFCLAIM if it states factual past or present information.:contentReference[oaicite:0]{index=0} The dataset combines analyst reports and earnings‑call transcripts, enabling research on the influence of numerical forecasts on market reactions.:contentReference[oaicite:1]{index=1}

---

Supported Tasks

Task	Objective	Typical Metrics
Numerical Claim Classification	Classify a sentence as INCLAIM or OUTOFCLAIM.	Accuracy, Precision, Recall, F1
Optimism Scoring	Produce a continuous optimism score derived from claim likelihood (research use‑case in paper).	Spearman / Pearson correlation with returns

---

Dataset Structure

configs:
  - config_name: main
    data_files:
      - split: train
        path: main/train-*
      - split: test
        path: main/test-*
dataset_info:
  features:
    - context: string   # Financial sentence
    - response: string  # Label: INCLAIM / OUTOFCLAIM
  splits:
    - name: train
      num_examples: 2_144
      num_bytes:    317_988
    - name: test
      num_examples:   537
      num_bytes:     79_011
  download_size: 208_114
  dataset_size: 396_999

Data Fields

Field	Type	Description
context	string	Sentence from an analyst report or earnings call.
response	string	INCLAIM or OUTOFCLAIM label.

---

Data Splits

Split	# Sentences	Portion
Train	2 144	80 %
Test	537	20 %
Total	2 681	100 %

---

Dataset Creation

• Source Collection – Analyst reports (Thomson Reuters) and quarterly earnings‑call transcripts for U.S. public firms (2010 – 2023).
• Sentence Filtering – Retained only sentences containing a financial term, numeric value, and currency/percentage symbol.
• Annotation – Weak‑supervision rules augmented with subject‑matter‑expert knowledge produced initial labels; a subset was manually validated.
• Quality Control – Manual spot‑checks achieved > 0.9 label accuracy, and noisy sentences were removed.

---

Usage

from datasets import load_dataset

numclaim = load_dataset("gtfintechlab/Numclaim")
sample = numclaim["train"][0]

print(sample["context"])
print(sample["response"])

License

Released under Creative Commons Attribution 4.0 International (CC BY 4.0).

Citation

@inproceedings{shah2024numclaim,
  title     = {Numerical Claim Detection in Finance: A New Financial Dataset, Weak-Supervision Model, and Market Analysis},
  author    = {Shah, Agam and Hiray, Arnav and Shah, Pratvi and Banerjee, Arkaprabha and Singh, Anushka and Eidnani, Dheeraj and Chava, Sahasra and Chaudhury, Bhaskar and Chava, Sudheer},
  booktitle = {Findings of the Association for Computational Linguistics},
  year      = {2024},
  url       = {https://arxiv.org/abs/2402.11728}
}