Political Tweet Sentiment & Framing

Overview

This project explored political communication on Twitter through an applied NLP pipeline for lexical analysis, sentiment profiling and emotion-based framing.

The work combined tweet collection, text preprocessing, word-frequency analysis, n-gram extraction, sentiment analysis and emotion lexicon matching to compare communication patterns across political actors and time periods.

The project should be read as an exploratory text-analytics prototype, not as a modern transformer-based sentiment system.

Problem

Political messages are not only informative; they also frame issues through specific vocabularies, emotional tones and repeated expressions.

The goal was to build a reproducible workflow for extracting these patterns from Twitter data and comparing political discourse through interpretable textual indicators.

Example 1 — Lexical and N-Gram Analysis

The pipeline tokenized tweet text, removed punctuation and stopwords, and computed frequent words, bigrams and trigrams.

tweet text
    ↓
tokenization
    ↓
cleaning and stopword removal
    ↓
word frequencies, bigrams and trigrams

This was used to identify recurrent terms and expressions associated with specific political actors or thematic periods.

Example 2 — Sentiment and Emotion Profiling

The project also used sentiment and emotion lexicons to summarize the emotional profile of political communication.

In particular, the workflow explored polarity scores and NRC-style emotion categories such as anger, fear, trust, joy and anticipation.

These outputs were used as descriptive indicators, not as definitive measures of political intent.

Technologies and Methods Used

• Python for tweet collection, preprocessing and exploratory analysis.
• R for text mining, visualization and exploratory sentiment workflows.
• GetOldTweets3 for historical tweet collection.
• pandas / NumPy for tabular data handling and intermediate structures.
• NLTK for tokenization, stopword removal and n-gram construction.
• TextBlob for exploratory polarity and subjectivity scoring.
• NRC Emotion Lexicon for emotion-category matching.
• tidytext / dplyr / ggplot2 for R-based text processing and visualization.
• matplotlib / seaborn for Python-based exploratory plots.

Implemented Elements

• Collection of tweet data for selected political accounts and periods.
• Cleaning of raw tweet text, links, punctuation and noisy tokens.
• Italian and English stopword filtering.
• Word-frequency analysis for political discourse comparison.
• Bigram and trigram extraction.
• Exploratory polarity and subjectivity scoring.
• Emotion profiling using lexicon-based matching.
• Bar plots and radar-style visualizations of lexical and emotional patterns.

Evaluation Status

The project produced descriptive NLP indicators rather than a supervised classification model.

The outputs are therefore useful for exploration and comparison, but they should not be interpreted as externally validated sentiment labels.

• Already present: lexical frequencies, n-grams, polarity scores and emotion profiles.
• Already present: visual comparison of political-text indicators.
• To be added: manually annotated validation set for sentiment or framing categories.
• To be added: robustness checks across time periods, topics and political actors.
• To be added: comparison with transformer-based language models for Italian text.

Methodological Note

Lexicon-based sentiment analysis is interpretable and useful for exploratory work, but it has important limitations.

It may fail with irony, sarcasm, negation, political slang, multilingual content and context-dependent meaning. For this reason, the results should be treated as descriptive signals rather than as ground-truth psychological or political measures.

Modern Extension

A modern version of this project would preserve the interpretable descriptive layer while adding contextual language models and stronger evaluation.

• Use transformer-based models for Italian sentiment and stance detection.
• Compare lexicon-based sentiment with supervised or zero-shot classifiers.
• Add topic modeling or embedding-based clustering to identify recurrent frames.
• Evaluate sentiment and framing outputs against human annotations.
• Track discourse dynamics over time and around political events.

Resources

Technical report in preparation.

Code available upon request.

Technical Context

• Mohammad & Turney, NRC Emotion Lexicon — relevant because the project used lexicon-based emotion categories for exploratory profiling.
• TextBlob — relevant because the prototype used polarity and subjectivity scores for exploratory sentiment analysis.
• Silge & Robinson, tidytext — relevant to the R-based text mining and visualization workflow.