Aspect-Based Medical Record Classification Using Large Language Model Guided Knowledge Graph

Traditional sentiment analysis approaches typically evaluate a text as a whole and assign it a single sentiment label, such as positive or negative. Although this method works well for many tasks, there are cases where it is more beneficial to understand sentiment related to specific aspects. To address this issue, Aspect-Based Sentiment Analysis (ABSA) focuses on analysing sentiment at the aspect level, treating it as a more detailed form of opinion mining. In this study, we proposed a method that initially identifies aspect terms as an extraction sub-task of anatomy terms by leveraging biomedical knowledge graphs. In the second subtask, we leverage well-known large language models to predict the sentiment polarities of these extracted aspect terms. The experimental results for each subtask demonstrate that the RaTE-NER-Deberta model yields the best performance in the anatomy aspect identification subtask, achieving precision, recall, and F1 scores of 65.385, 64.151, and 64.762, respectively. After identifying anatomical entities in the input texts using this model, we proceed with the classification task. The deberta-v3-base-absa-v1.1 model, a specialized version for aspect-based sentiment analysis, delivers the highest results, with a precision of 91.38, recall of 80.30, and an F1 score of 85.48.