Political Discourse Dataset

Overview

This dataset complements the USE24-XD dataset and contains social media posts annotated with political candidate affiliation, temporal context, and topic tokens derived from LDA preprocessing.

It is designed for research on:

• Political discourse analysis
• Temporal dynamics of narratives
• Topic modeling and NLP tasks

---

Data Schema

Each row represents a single post with the following fields:

Column	Type	Description
id	int/string	Unique identifier for the post
candidate_fan	string	Candidate affiliation label: Trump, Neutral, or Biden_Kamala
user_state	string	Inferred U.S. state from user-provided location
sentiment_score	float	Continuous sentiment score in the range [-1, 1]
sentiment_label	string	Sentiment category: positive, negative, or neutral
hate_count	int	Number of detected hate-related keywords
hate_flag	int/bool	Binary indicator for hate-related content
misinfo_count	int	Number of detected misinformation-related keywords
misinfo_flag	int/bool	Binary indicator for misinformation-related content
tweet_date	date/string	Date when the post was created
tweet_hour	int	Hour of day when the post was created, from 0 to 23
tweet_dayofweek	int/string	Day of week when the post was created
temporal_period	string	Time period relative to the election or event
engagement_raw	int/float	Total engagement count
engagement_rate	float	Engagement normalized by impressions
engagement_rate_log	float	Log-transformed engagement rate
sensitive_flag	int/bool	Binary indicator for platform-flagged sensitive content
lda_tokens	list[string]	Preprocessed tokens used for topic modeling

Example

👉 Checkout preview: annotation_data_sample100.csv

Sample Rows

id	candidate_affiliation	user_state	sentiment_score	sentiment_label	hate_count	hate_flag	misinfo_count	misinfo_flag	tweet_date	tweet_hour	tweet_dayofweek	engagement_raw	engagement_rate	engagement_rate_log	sensitive_flag	lda_tokens
1882403261095043491	Trump	Non_US	-0.8658	negative	3	1	0	0	2025-01-23	12	3	0	0	0	0	['moron', 'get', 'ass', 'kick', 'election', ...]
1932964629195731097	Trump	CA	-0.9482	negative	3	1	1	1	2025-06-12	0	3	2	0.06896551724	0.0666913745	0	['evidence', 'emerge', 'prove', 'wrongdoing', ...]
1886480963905183809	Trump	Non_US	-0.9779	negative	3	1	0	0	2025-02-03	18	0	0	0	0	1	['mean', 'trump', 'deport', 'well', ...]

Stance Detection Experiments

This repository contains the notebook stance_detection_experiment.ipynb, which implements the full pipeline for political stance detection on social media data. The notebook covers preprocessing, feature engineering, model training, and evaluation across multiple machine learning approaches.

---

Overview

The notebook provides a unified experimental framework to compare:

• Classical machine learning models
• Neural models using sentence embeddings
• Transformer-based models

All models are evaluated under a consistent preprocessing and evaluation pipeline.

---

Pipeline Components

1. Data Preprocessing

The notebook applies a multi-stage preprocessing pipeline:

• Text normalization and cleaning
• Removal of URLs, mentions, and noise
• Tokenization and lemmatization
• Hashtag and emoji handling
• Filtering short or empty posts

---

2. Feature Engineering

The following features are constructed:

Text Features

• TF-IDF representations (unigrams + bigrams)
• Sentence embeddings using SBERT (all-mpnet-base-v2)

Sentiment Features

• Sentiment score in [-1, 1]
• Sentiment label (positive, negative, neutral)

Hate and Misinformation Signals

• Keyword-based counts and binary flags

Temporal Features

• Hour of day
• Day of week
• Event-based temporal period

Engagement Features

• Raw engagement counts
• Normalized engagement rate
• Log-transformed engagement rate

Metadata

• User location (state-level)
• Platform sensitivity flag

Topic Modeling

• LDA token features for thematic analysis

---

3. Models Implemented

The notebook includes the following models:

Classical Machine Learning

• Logistic Regression (TF-IDF features)
• Linear Support Vector Machine (SVM)
• Histogram Gradient Boosting (HGB with SVD)

Neural Model

• SBERT (all-mpnet-base-v2) embeddings + MLP classifier

Transformer Model

• BERTweet (vinai/bertweet-base) fine-tuned for stance classification

---

4. Training Setup

• Stratified data split:

  • 64% training
  • 16% validation
  • 20% test

• Early stopping based on validation macro-F1

• Hyperparameter tuning for all models

---

5. Evaluation

Models are evaluated using:

• Macro-F1 score (primary metric)
• Accuracy
• Confusion matrices for error analysis

---

How to Run

Open the notebook:

jupyter notebook stance_detection_experiment.ipynb