Chinese Sentiment Analysis Fund Direction

🚀 Chinese Sentiment Analysis Model for Fund Direction

This is a Chinese sentiment analysis model for the fund sector, trained with over 100,000 data points. It has shown good performance in initial tests. The negative texts have been specially processed, while the neutral ones might not be very accurate.

📚 Documentation

Model Evaluation

• Validation Set Accuracy: 0.9382193411826961
• Validation Set Classification Report:

              precision    recall  f1-score   support

    negative       0.93      0.95      0.94      3785
    positive       0.95      0.96      0.95      6919
     neutral       0.93      0.89      0.91      4414

    accuracy                           0.94     15118
   macro avg       0.94      0.93      0.93     15118
weighted avg       0.94      0.94      0.94     15118

Return Value Explanation

• 0: 'negative', 1: 'positive', 2: 'neutral'

💻 Usage Examples

Basic Usage

import sys
import re
import torch
from transformers import BertTokenizer, BertForSequenceClassification
from torch.nn.functional import softmax

# Set to use CPU or CUDA
device = 'cuda' if torch.cuda.is_available() else 'cpu'

# Load the pre-saved model and tokenizer
model = BertForSequenceClassification.from_pretrained('sanshizhang/Chinese-Sentiment-Analysis-Fund-Direction')
tokenizer = BertTokenizer.from_pretrained('sanshizhang/Chinese-Sentiment-Analysis-Fund-Direction')

# Ensure the model is on the correct device
model = model.to(device)
model.eval()  # Set the model to evaluation mode

# Function definition: Make predictions and return prediction probabilities
def predict_sentiment(text):
    # Encode text data
    encoding = tokenizer.encode_plus(
        text,
        max_length=512,
        add_special_tokens=True,
        return_token_type_ids=False,
        padding='max_length',  # Modify here
        return_attention_mask=True,
        return_tensors='pt',
        truncation=True
    )
    
    # Extract the corresponding encoding of the input
    input_ids = encoding['input_ids'].to(device)
    attention_mask = encoding['attention_mask'].to(device)
    
    # Do not calculate gradients
    with torch.no_grad():
        # Generate logits for sentiment prediction
        outputs = model(input_ids=input_ids, attention_mask=attention_mask)
    
    # Use softmax to convert logits to probabilities
    probs = softmax(outputs.logits, dim=1)

    # Return probabilities and the predicted class
    return probs, torch.argmax(probs, dim=1).cpu().numpy()[0]

# Get text from command-line arguments, merge and clean special characters
arguments = sys.argv[1:]  # Ignore the script name
text = ' '.join(arguments)  # Merge into a single string
text = re.sub(r"[^\u4e00-\u9fff\d.a-zA-Z%+\-。！？，、；：（）【】《》“”‘’]", '', text)  # Remove special characters

#print(f"Passed text is: {text}")
# Make predictions
probabilities, prediction = predict_sentiment(text)

sentiment_labels = {0: 'negative', 1: 'positive', 2: 'neutral'}

# Print the predicted sentiment and its probability
predicted_sentiment = sentiment_labels[prediction]
print(f"Predicted sentiment: {predicted_sentiment},Probability:{probabilities[0][prediction].item()}")
#print(f"Probability: {probabilities[0][prediction].item()}")

📄 License

This project is licensed under the Apache-2.0 license.