Emotion-Focused Analysis of Stock Tweets

This repository contains the code to perform multi-class emotion classification and topic modeling on stock-related tweets using various NLP techniques.

Link to the StockEmotion Dataset.

Project Summary

We investigate:

• Emotion classification using TF-IDF, Word2Vec, and contextual embeddings (BERTweet, Distil-RoBERTa)
• Topic modeling with LDA and BERTopic
• Integration of emoji features and sentiment lexicons (VADER, NRC, Bing Liu)

For full details, see the project report.

Notebooks Overview

Notebook	Description
dataset_exploration.ipynb	Exploratory Data Analysis (EDA)
TM_preproc_FE_classification.ipynb	Preprocessing, feature extraction, training of classifiers
word2vec_multi.ipynb	Word2Vec training and downstream classification
BERT_embeddings_and_classifier.ipynb	Embedding generation (BERTweet, Distil-RoBERTa) + classifier
topic_modeling_LDA.ipynb	LDA-based topic modeling
topic_modeling_BERTopic.ipynb	BERTopic-based topic modeling

Note: Pre-trained models (e.g., Word2Vec) and large files are excluded from the repo.

Environment Setup

This project uses Python 3.12. It's recommended to create a virtual environment to manage dependencies.

Create and activate a virtual environment (venv)

Unix/macOS:

python3.12 -m venv venv
source venv/bin/activate

Windows:

py -3.12 -m venv venv
venv\Scripts\activate

Install dependencies

Start by upgrading pip:

pip install --upgrade pip

Install required packages manually:

pip install bertopic gensim nltk spacy emoji vaderSentiment

Additional packages may be required based on notebook usage (e.g., scikit-learn, matplotlib, pandas, umap-learn).

Set up NLTK and spaCy

After installing:

import nltk
nltk.download('stopwords')
nltk.download('punkt')

import spacy
spacy.cli.download("en_core_web_sm")

External Files Needed

• Lexicons: Bing Liu and NRC Emotion Lexicons (update paths accordingly).
• Processed CSV: Required before running topic_modeling_BERTopic.ipynb.
• Pre-trained Word2Vec: Must be trained separately or loaded from local storage.

Findings

• Best classifier: XGBoost with Bigram TF-IDF (macro-F1 ≈ 0.32)
• Topic modeling coherence: ~0.3, indicating challenges due to tweet brevity/noise
• Emoji features proved critical in both classification and topic modeling

Authors

• Robin Smith
• Sergio Verga

Università degli Studi di Milano-Bicocca

References

• BERTweet
• VADER
• NRC Emotion Lexicon
• Bing Liu's Lexicon

For methodology and evaluation, see: TM_report_Smith_Verga.pdf.