---
language:
- en
license: mit
task_categories:
- information-extraction
- text-classification
- token-classification
tags:
- financial-nlp
- earnings-calls
- sec-filings
- kpi-extraction
- large-language-models
pretty_name: Effective Performance Measurement (ECB & SECB)
---

# Effective Performance Measurement: KPI Extraction Datasets

This dataset repository accompanies the ACL 2026 (Industry Track) paper: **"Effective Performance Measurement: Challenges and Opportunities in KPI Extraction from Earnings Calls"**. 

It contains three novel benchmarks and a prediction set designed to evaluate the extraction of Key Performance Indicators (KPIs) from unstructured financial texts, specifically comparing highly regulated SEC filings to conversational earnings calls.

🔗 **Associated GitHub Repository:** [AAU-NLP/effective-performance-measurement](https://github.com/AAU-NLP/effective-performance-measurement)

---

## 📊 Dataset Overview

This repository includes four data files, covering data from 20 S&P 500 companies between 2023 and 2024.

### 1. SEC Filings Benchmark (`SECB.json`)
* **Description:** Context-rich text chunks extracted from SEC filings (10-K, 10-Q). This dataset serves as a baseline to test models trained on highly structured, templated financial data.
* **Size:** 40,661 chunks.
* **Annotations:** 77,677 regex-labeled entities.

### 2. Earnings Call Benchmark (`ECB.json`)
* **Description:** Raw, unannotated conversational text chunks extracted from corporate earnings calls. This represents the challenging, unstructured domain shift.
* **Size:** 10,477 chunks.

### 3. Annotated Earnings Call Benchmark (`ECB-A.json`)
* **Description:** An expert-annotated subset of the ECB dataset used for evaluating Large Language Model (LLM) extraction and in-context learning techniques.
* **Size:** 587 chunks.
* **Annotations:** 2,460 entities and 934 relational groups.

### 4. Longitudinal KPI Tracking (`gold_standard_traceable.jsonl`)
* **Description:** A dataset containing post-hoc semantic clusterings of KPIs to track emergent metrics across multiple quarters.
* **Size:** 1,323 traced entity/KPI rows.

---

## 💻 How to Load the Data

You can easily load this data using the Hugging Face `datasets` library, or by downloading the JSON files directly.

```python
from datasets import load_dataset

# Load the entire repository
dataset = load_dataset("AAU-NLP/effective-performance-measurement")

# Alternatively, download specific JSON files if you just want one benchmark
# e.g., wget [https://huggingface.co/datasets/AAU-NLP/effective-performance-measurement/resolve/main/ECB-A.json](https://huggingface.co/datasets/AAU-NLP/effective-performance-measurement/resolve/main/ECB-A.json)
```