---
license: mit
language:
- en
tags:
- financial-nlp
- sentiment-analysis
- topic-classification
- multitask-learning
- bert
- financial-news
library_name: transformers
pipeline_tag: text-classification
datasets:
- financial-news
metrics:
- accuracy
- f1
- precision
- recall
---

# Multi-Task BERT for Financial News Topic Classification and Sentiment Analysis

## Model Description

This model is a multi-task BERT-based architecture designed to simultaneously perform topic classification and sentiment analysis on financial news text. The model leverages shared representations to improve performance on both tasks through multi-task learning.

## Model Details

- **Model Type**: Multi-task BERT for text classification
- **Language**: English
- **License**: MIT
- **Tasks**: 
  - Topic Classification (financial news categories)
  - Sentiment Analysis (positive, negative, neutral)

## Intended Uses

### Direct Use

This model can be used for:
- Analyzing sentiment in financial news articles
- Classifying financial news into relevant topics/categories
- Automated content analysis for financial research
- Risk assessment based on news sentiment

### Downstream Use

The model can be fine-tuned for:
- Specific financial domains (stocks, forex, commodities)
- Custom topic taxonomies
- Different sentiment granularities

## How to Use

```python
import torch
import pickle
from transformers import AutoTokenizer, AutoModel

# Load the model
with open('multitask_bert_model.pkl', 'rb') as f:
    model = pickle.load(f)

# Load tokenizer (adjust model name as needed)
tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')

# Example usage
text = "Apple stock rises 5% after strong quarterly earnings report"
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=512)

# Get predictions (adjust based on your model's output format)
with torch.no_grad():
    outputs = model(**inputs)
    # Process outputs for topic and sentiment predictions
```

## Training Data

The model was trained on financial news data for multi-task learning. The training involved:
- Topic classification task
- Sentiment analysis task
- Joint optimization with shared BERT representations

## Training Procedure

### Training Hyperparameters

- **Training regime**: Multi-task learning with shared encoder
- **Model variants**: 
  - `multitask_bert_model.pkl`: Base model
  - `multitask_bert_model_weight.pth`: Weighted version
  - `multitask_bert_model_imbalanced.pth`: Version trained on imbalanced data

### Training Details

The model uses a shared BERT encoder with task-specific classification heads for topic classification and sentiment analysis. The multi-task approach allows the model to learn shared representations that benefit both tasks.

## Evaluation

### Testing Data & Metrics

The model should be evaluated on:
- **Topic Classification**: Accuracy, F1-score, Precision, Recall
- **Sentiment Analysis**: Accuracy, F1-score, Precision, Recall

### Results

[Add your evaluation results here]

| Task | Metric | Score |
|------|--------|-------|
| Topic Classification | Accuracy | 0.76 |
| Sentiment Analysis | Accuracy | 0.87 |

## Limitations and Bias

### Limitations

- Performance may vary on financial news from different time periods
- Model may not generalize well to non-financial text
- Limited to English language text
- Performance depends on the quality and diversity of training data

### Bias Considerations

- Model may reflect biases present in financial news training data
- Sentiment predictions may be influenced by market conditions during training
- Topic classifications may favor certain financial sectors represented in training data

## Technical Specifications

### Model Architecture

- **Base Model**: BERT
- **Architecture**: Multi-task learning with shared encoder