Classification of full-length news articles
| task | representation/model | metric |
|---|---|---|
| binary topic classification | RoBERTa-large fine-tuned | 93% accuracy, 91% recall (test size = 0.2, 2890) |
| 3-class sentiment classification | RoBERTa-large fine-tuned | 75% accuracy (test size = 0.2, 1530) |
| class | # |
|---|---|
| financial | 7647 |
| non-financial | 6799 |
| classification | # |
|---|---|
| neg | 2287 |
| neu* | 2544 |
| pos | 2816 |
*a major source of uncertainty and error(during manual classification and modeling),
LLMs
https://magazine.sebastianraschka.com/p/understanding-large-language-models
Huggingface (Transformers)
https://huggingface.co/docs/transformers/index
https://huggingface.co/docs/transformers/tasks/sequence_classification
https://huggingface.co/docs/transformers/training#train-with-pytorch-trainer
https://huggingface.co/docs/transformers/v4.30.0/en/main_classes/trainer#transformers.Trainer
Scikit-Learn
https://scikit-learn.org/stable/user_guide.html
Evaluation of Sentiment Analysis in Finance: From Lexicons to Transformers
https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9142175
FinBERT: Financial Sentiment Analysis with Pre-trained Language Models
https://arxiv.org/pdf/1908.10063.pdf
How to Fine-Tune BERT for Text Classification?
https://arxiv.org/pdf/1905.05583.pdf
Financial Sentiment Analysis:
An Investigation into Common Mistakes and Silver Bullets
https://aclanthology.org/2020.coling-main.85.pdf