Extracting Fine-Grained Sentiment Features about Library Services from Reader Feedback Text Using RoBERTa

Abstract

Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa) model, fine-tuned for domain adaptation with library domain data, and introducing the Aspect-Based Sentiment Analysis (ABSA) framework to extract and classify fine-grained sentiment related to collection services in reader feedback. Based on annotation and experiments with 23,684 open reader feedback texts from 12 university libraries in China over the past three years, a fine-grained annotation system is constructed, encompassing six collection service dimensions, three types of sentiment polarity, including positive, neutral, and negative, and three levels of intensity. The results show that the domain-fine-tuned RoBERTa model achieved an F1 score of 86.7% in aspect recognition and an accuracy of 89.3% in sentiment polarity classification. The aspect recognition F1 score is significantly better than baseline models such as Bidirectional Encoder Representations from Transformers (BERT) by 5.2% and Text Convolutional Neural Network (TextCNN) by 11.6%. Among the evaluated dimensions, access to electronic resources and speed of obtaining books are the two dimensions with the highest concentration of negative sentiment, accounting for 31.4% and 27.8% of the total negative feedback, respectively. This research method improves the reliability and interpretability of sentiment analysis in professional service scenarios and provides a data-driven reference for optimizing digital resource services for engineering disciplines involving electromagnetic theory, antenna technology, and propagation research.