Sentiment analysis with Hierarchical Attention Topic Aware Hierarchical Multi-Attention Network using citation sentences

Abstract

Sentiment analysis plays a critical role in identifying the underlying opinions, emotions and attitudes conveyed in textual data. In the context of academic literature, accurately interpreting the sentiment behind citations is essential for understanding the citing author’s intent and the overall impact of cited works. However, existing sentiment analysis methods often struggle to capture the nuanced and domain-specific language of academic writing. These methods typically rely on shallow linguistic features or general-purpose models, resulting in limited contextual understanding and reduced accuracy in detecting subtle or implicit sentiments. To solve these limitations, a Hierarchical Attention Network with Topic aware Hierarchical Multi-Attention Network (HAN-T-HMAN) model is devised for analyzing sentiment utilizing citation sentences. Initially, the citation sentences are tokenized by Bidirectional Encoder Representations from Transformers (BERT) tokenization and Aspect Term Extraction (ATE) is performed. The features are extracted at the feature extraction stage. At last, HAN-T-HMAN performs sentiment analysis, which is formed by incorporating Hierarchical Attention Network (HAN) and Topic aware hierarchical multi-attention network (T-HMAN). The proposed HAN-T-HMAN model has attained superior values for metrics, such as precision, recall, F1-Score, and accuracy with values of 92.99%, 93.15%, 93.07%, and 94.85%.