Sentiment Classification for Financial Texts Based on Deep Learning.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Financial Technology (FinTech)

Background:

• Sentiment classification for financial texts is crucial for market prediction and crisis identification.
• Deep learning applications in NLP are expanding, but financial text sentiment analysis is hindered by a scarcity of labeled data.
• Existing methods struggle with the limited availability of domain-specific labeled financial datasets.

Purpose of the Study:

• To propose a domain-adaptation method for financial text sentiment classification.
• To address the challenge of limited labeled samples in the financial domain.
• To enhance the accuracy of sentiment classification in financial contexts using transfer learning.

Main Methods:

• A novel cross-domain transfer learning method is introduced.
• A domain classification subnetwork and loss function are integrated into a neural network.
• The model is trained using labeled source domain data and unlabeled target domain financial data.

Main Results:

• The proposed method achieved classification accuracy rates of 65.0% (Books), 61.2% (DVDs), 61.6% (Electronics), and 66.3% (Kitchen Appliances) using Amazon reviews as the source domain.
• Compared to non-transfer learning approaches, accuracy improved by 11.0%, 7.6%, 11.4%, and 13.4% across the tested domains.
• The model demonstrated effective adaptation to the target financial domain while performing the classification task.

Conclusions:

• The proposed domain-adaptation method effectively enhances financial text sentiment classification accuracy.
• Cross-domain transfer learning offers a viable solution to the problem of limited labeled data in specialized domains like finance.
• This approach holds significant potential for improving financial market prediction and analysis through NLP.