Financial Text Summarization Project
📍 Live App: Try it on Hugging Face
Problem & Motivation
Financial analysts often face information overload when reviewing long earnings reports, transcripts, and market commentary.
This project builds a transformer-based system to:
- Recognize speech from meetings/calls (Wav2Vec)
- Summarize long financial text (BART)
- Analyze financial sentiment (FinBERT)
Goal: Deliver fast, digestible, and actionable insights from unstructured financial data.
Model Architecture
| Task | Model Used |
|---|---|
| Speech Recognition | facebook/wav2vec2-base-960h |
| Text Summarization | knkarthick/MEETING_SUMMARY (BART) |
| Sentiment Analysis | yiyanghkust/finbert-tone (BERT variant) |
These models are orchestrated in a Gradio UI, enabling real-time interaction.
Workflow
- Speech-to-Text: Users can record/upload earnings call audio
- Text Summarization: BART reduces raw text to 3–4 key financial insights
- Tone Classification: BERT classifies sentiment as Positive, Neutral, or Negative
This pipeline empowers analysts to quickly identify risks, sentiment, and trends.
Technical Highlights
- Transformer Models: Leveraged transfer learning from pretrained BERT, BART, and Wav2Vec2
- Financial Domain Fine-Tuning: Enhanced summarization accuracy with finance-specific datasets
- Web Deployment: Hosted on Hugging Face Spaces using Gradio for rapid access
Business Value
- Saves analysts hours of manual review of financial disclosures
- Supports investment decision-making with real-time sentiment
- Powers AI-driven insights for traders, PMs, and hedge funds
- Framework is scalable to global markets with future multilingual support
Future Enhancements
- Multilingual report support (e.g. earnings calls in Chinese, Japanese)
- Real-time news stream summarization
- Integration with automated trading platforms
- Continual fine-tuning on latest market data for relevance
📎 PDF Report: Download Here
🔗 Live Demo: https://huggingface.co/spaces/Vickiiiyippp/financial_text_summarization