講演情報
[4O1-IS-2a-01]Financial Sentiment Domain Adaptation via Word Polarity Adaptation with Interpretable Neural Networks
〇Tomoki Ito1 (1. NICT)
regular
キーワード：
Interpretable Neural Network、Financial Sentiment Analysis、Domain Adaptation
Large Language Models (LLMs) are widely deployed in real-world applications, where auxiliary classifiers are used to control or filter generated outputs. In high-stakes domains such as finance and legal services, such systems must ensure both interpretability and computational efficiency. Conventional fine-tuning methods update many parameters, increasing adaptation cost and obscuring how individual words influence decisions.
We propose an interpretable and efficient neural network. We extend the Sentiment Shift Neural Network (SSNN) with pretrained BERT attention, resulting in BERT-SSNN, which preserves word-level polarity visualization and sentiment shift analysis while improving predictive performance.
We further introduce Word-level Polarity Adaptation, which updates only the Word-level Original Sentiment Layer while freezing context-related layers. This design reduces training cost and maintains transparency by directly linking parameter updates to word-level polarity changes. Experiments demonstrate competitive performance with substantially lower adaptation cost.
We propose an interpretable and efficient neural network. We extend the Sentiment Shift Neural Network (SSNN) with pretrained BERT attention, resulting in BERT-SSNN, which preserves word-level polarity visualization and sentiment shift analysis while improving predictive performance.
We further introduce Word-level Polarity Adaptation, which updates only the Word-level Original Sentiment Layer while freezing context-related layers. This design reduces training cost and maintains transparency by directly linking parameter updates to word-level polarity changes. Experiments demonstrate competitive performance with substantially lower adaptation cost.