Financial PhraseBank Sentiment Analysis & Topic Modeling with LLMs

Overview

This project provides a full pipeline for analyzing financial news sentences using state-of-the-art NLP techniques. It includes:

• Data cleaning and exploratory analysis
• Topic modeling with LDA
• Sentiment classification using FinBERT and open-source LLMs
• Retrieval-Augmented Generation (RAG) for improved LLM performance
• Fine-tuning FinBERT for domain-specific sentiment analysis

Steps

1. Preprocessing & EDA:

   • Load and clean the Financial PhraseBank dataset.
   • Analyze sentiment distribution and text statistics.

2. Topic Modeling:

   • Apply LDA to discover and interpret latent topics in financial text.

3. Sentiment Analysis:

   • Use FinBERT for sentiment prediction.
   • Evaluate with accuracy, F1-score, and confusion matrix.

4. Local LLM Sentiment Analysis:

   • Run Mistral-7B or similar LLMs for zero-shot/few-shot sentiment classification.
   • Compare with FinBERT results.

5. RAG:

   • Retrieve similar examples using FAISS and sentence-transformers.
   • Augment LLM prompts for better sentiment prediction.

6. Fine-Tuning:

   • Fine-tune FinBERT on the dataset.
   • Visualize and evaluate the improved model.

Outputs

• Preprocessed data and topic assignments
• Sentiment predictions from FinBERT, LLM, and RAG
• Evaluation metrics and visualizations
• Fine-tuned FinBERT model

Requirements

• Python (with Jupyter Notebook)
• pandas, numpy, matplotlib, seaborn, nltk, scikit-learn
• torch, transformers, sentence-transformers, faiss

Usage

Open DLP_Assignment_3.ipynb and run the cells step by step. Follow the comments for guidance.