pauladroghoff/manipulative-language-detection

Manipulative Language Detection Dataset

This dataset contains annotated text examples for detecting manipulative language at both sentence and dialogue levels.

Dataset Description

The Manipulative Language Detection Dataset is designed to help train and evaluate transformer-based models in identifying manipulative language patterns. The dataset consists of two complementary components:

1. Sentence-level data: Individual sentences labeled as manipulative (1) or non-manipulative (0)
2. Dialogue-level data: Conversational exchanges with annotations for manipulation techniques, victim vulnerabilities, and context

The dataset is sourced from various dialogues, including movie scripts and other conversational contexts. Each entry is thoroughly annotated for manipulation attributes.

This work is aligned with recent research on mental manipulation detection, such as the MentalManip dataset (Wang et al., 2024).

Data Format

Sentence-Level Data

Each entry contains:

• Inner ID
• Unique ID
• Sentence text
• Binary manipulation label (1=manipulative, 0=non-manipulative)
• Original context (dialogue source)
• Movie name
• Annotator agreement metrics
• Manipulation technique categorization (persuasion, intimidation, seduction, etc.)
• Victim/vulnerability annotations
• Confidence scores

Dialogue-Level Data

Each entry contains:

• Inner ID
• Unique ID
• Dialogue exchange with speaker identification
• Manipulation classification (binary)
• Movie Name
• Annotator agreement metrics
• Manipulation technique categorization (persuasion, intimidation, seduction, etc.)
• Victim/vulnerability annotations
• Confidence scores

Manipulation Techniques

The dataset identifies several manipulation techniques, including:

• Persuasion or Seduction
• Accusation
• Denial
• Evasion
• Feigning Innocence
• Rationalization
• Playing the Victim Role
• Playing the Servant Role
• Shaming or Belittlement
• Intimidation
• Brandishing Anger

Targeted Vulnerability

The dataset identifies several vulnerability targets, including:

• Over-responsibility
• Over-intellectualization
• Naivete
• Low self-esteem
• Dependency

Usage

This dataset is designed for training transformer-based models to detect manipulative language. Researchers can use it to:

1. Train binary classifiers at the sentence level and/or dialogue level
2. Develop more sophisticated models that identify specific manipulation techniques
3. Study the contextual nature of manipulation in dialogues
4. Evaluate models' performance across different manipulation strategies

Evaluation Metrics

Models can be evaluated using standard metrics including:

• Accuracy
• F1 score
• Precision
• Recall
• Balanced accuracy

Research has shown that current language models, even advanced ones like GPT-4, struggle with correctly identifying and classifying manipulative language patterns, highlighting the importance of specialized datasets and models for this task.

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("pauladroghoff/manipulative-language-detection")

# Access sentence-level data
sentence_data = dataset["sentence_level"]

# Access dialogue-level data
dialogue_data = dataset["dialogue_level"]