Fine-Grained Feature Extraction in Key Sentence Selection for Explainable Sentiment Classification Using BERT and CNN

Abstract

Online product reviews provide valuable insights into customer sentiment toward products; however, they often contain multiple sentences with redundant and non-essential content, making it harder to extract critical information. Identifying key sentences that directly impact sentiment along with fine-grained feature extraction is essential for reliable multi-class sentiment classification with greater explainability. Existing studies on document-level sentiment analysis with sentence-wise learning often overlook important local feature extraction at both the sentence and document levels for multi-class sentiment classification, limiting the selection of key sentences with greater transparency. These models rely on sequential processing of text rather than parallel processing, which can lead to missed important fine-grained features and long-range dependencies spread across sentences. Additionally, sequential models place more emphasis on the latter part of the text, assuming that the emotional aspects are concentrated at the end of the review. To address this issue, we introduce a novel Bidirectional Encoder Representations from Transformers (BERT) based dilated sentence attention deep learning model (Bert_DSENT_Att) for important fine-grained local feature extraction by capturing long-range dependencies within sentences, thereby enabling the selection of key sentences that directly impact multi-class sentiment classification. First, the BERT pre-trained language model captures global features through a feature vector representation. Next, a multi-layered Convolutional Neural Network (CNN) with varying kernels and dilation rates extracts long range local features at the sentence level. We employ within-sentence attention and between-sentence attention to extract important fine-grained features from the hierarchical representation. The experimental results show that our model outperforms seven baselines across three datasets in terms of accuracy, precision, recall, and F1 measure. A thorough ablation study demonstrated the contribution of each component to overall performance.