📦 LegitimNarrate Dataset

The LegitimNarrate dataset is a structured collection of crowdfunding project descriptions annotated for legitimation mechanisms. It is designed for sequential sentence classification and legitimation analysis in entrepreneurial narratives. The dataset is split into three files located in the data/ directory:

📂 Dataset Files

The dataset is organized in the data/ directory as follows:

├── legitimacy_train_.json
├── legitimacy_val_.json 
└── legitimacy_test_.json 

---

📚 Dataset Structure

Each file is in JSON Lines format, where each line represents a single crowdfunding project with its corresponding metadata and annotated labels.

📌 Example

{
  "project_id": 1,
  "sentences": [
    "Introduction : Octav is high quality speaker to assemble by yourself.",
    "It is entirely customizable and it allows you to create a product which suits you.",
    "It is also an upgradeable and repairable product, allowing you to extend its life cycle and consequently to have a controlled and eco-friendly way of buying.",
    ...
  ],
  "labels": [0, 0, 1, 0, ...],
  "dominant_mechanism": [0, 0, 1, 0, ...],
  "second_mechanism": [1, 1, 2, 3, ...],
  "legitim_score": 1.442695041,
  "status": "failed",
  "technology-category": "sound"
}

---

🏷 Field Descriptions

Field	Type	Description
project_id	int	Unique identifier for each crowdfunding project.
sentences	List[str]	A list of sentences from the project’s description.
labels	List[int]	Sentence-level binary labels: 1 if the sentence contains a legitimation mechanism, 0 otherwise.
dominant_mechanism	List[int]	Integer-encoded dominant legitimation mechanism per sentence. Values: 0 (none), 1 (identity), 2 (associative), 3 (organizational).
second_mechanism	List[int or null]	Second legitimation mechanism per sentence, if present. Otherwise null.
legitim_score	float	Overall legitimacy score of the project, calculated from the frequency of mechanisms normalized by sentence count.
status	str	Outcome of the campaign: successful , failed or canceled.
technology-category	str	The technological domain or category of the project (e.g., "sound", "health", "software").

---

🧪 Dataset Splits

The dataset is divided into three subsets and stored in the data/ directory:

File Name	Description
legitimacy_train_.json	Training set (~80% of the dataset)
legitimacy_val_.json	Validation set (~10%)
legitimacy_test_.json	Test set (~10%)

---

🆕 Augmented Data

In addition to the original dataset files, the data/ directory also includes an augmentation file generated using Bing Chat with varied tones (e.g., professional, casual, informative). This file is intended to support research in data augmentation, style variation, and robustness evaluation.

📄 File: augmentation.csv

Column	Description
augmented	The reformulated sentence generated in a different tone or style.
original	The original sentence from the dataset.
category	Label indicating whether the original sentence expresses legitimation mechanism (e.g., 1) or not (0).

🧪 Example Row

augmented: The product lifecycle has been extended to support an environmentally responsible approach. original: The product life cycle is extended, allowing you to enter into an eco-responsible approach. category: 1

🔍 Applications

The LegitimNarrate dataset supports a wide range of research and development applications, including:

• Entrepreneurial Legitimation Analysis
  Study how legitimation mechanisms are used in crowdfunding narratives and how they relate to campaign success.

• Sequential Sentence Classification
  Train models that account for the order and context of sentences to detect legitimation mechanisms.

• Discourse-level Narrative Understanding
  Explore how coherent and persuasive storytelling unfolds across multiple sentences in project descriptions.

• Imbalanced & Few-shot Learning
  Benchmark and develop methods to handle class imbalance, especially in detecting underrepresented mechanisms.

• AI for Social Impact and Communication Studies
  Analyze how entrepreneurs construct legitimacy through language, identity, and associative strategies in digital platforms.

• Style-based Data Augmentation
  Use the augmentation.csv file to expand training data with reformulated sentences across diverse tones (e.g., professional, casual, informative), improving generalization and robustness.

• Tone-aware Sentence Classification
  Investigate how different writing styles affect the perception or detection of legitimation strategies in narratives.

• Paraphrase Detection and Semantic Similarity
  Use aligned original and augmented pairs for developing and testing models that detect meaning preservation across stylistic shifts.

• Generated Text Detection
  Train models to distinguish between human-written and machine-generated sentences, using the augmentation.csv file as a labeled dataset for generation detection tasks.