Fintext_NLP

Fintext_NLP is a collection of applied projects exploring natural language processing in financial contexts. The goal is to evaluate how transformer-based models can be adapted, fine-tuned, or extended for tasks like sentiment analysis, domain adaptation, and transfer learning—specifically within the language of markets, news, and financial commentary.

Projects are modular, experiment-driven, and focused on bridging the gap between general-purpose language models and domain-specific performance in finance.

Navigate to: (Main Notebook)

---

Table of Contents

1. Introduction and Objective

2. Exploratory Data Analysis (EDA)

3. Training Classifier Head Only

   • 3.1 Experiment 1: Model Preformance Across Varying Subsets (Subset Fine-Tuning Experiments)
   • 3.2 Experiment 2: Addressing Class Imbalance (Class Imbalance Experiments)
   • 3.3 Aside: Working with TFDS - TensorFlow Datasets (TFDS Demonstration)

4. Supervised Fine-Tuning with Fixed Classifier Head

   • 4.1 Load Trained Model, Freeze Classifier Head, Unfreeze Encoder

5. Full Fine-Tuning of Entire Model (Pre-Trained + Classifier Head)

   • 5.1 Unfreezing
   • 5.2 Experiment 3: Detailed Fine Tuning (Hyperparameter Fine-Tuning Experiments)
   • 5.3 Experiment 4: Fine-Tuning with Custom Head Architecture

6. Low Rank Adaptation (LoRA)

   • LoRA Notebook

7. Evaluating Perfomance on Financial PhraseBank Variants

   • Alternate Flavors Experiments

8. Evaluating Alternative Hugging Face Models on Financial PhraseBank

   • Alternate HuggingFace Models

9. In-Context Learning Experiments

   • In-Context Learning Notebook

10. NLP for Commodities Trading

    • #OOTT Sentiment Analysis

11. Conclusion

---

Summary

This repository serves as a modular framework for testing and comparing various adaptation strategies for transformer models in financial NLP. Techniques are extensible to other datasets and domains.