A fine-grained sentiment analysis model based on multi-task learning
Xin Nan Fan, Zhonglin Zhang · 2024
Fine-grained sentiment analysis based on textual data is currently a prominent research topic in natural language processing. The objective of this analysis is to forecast various dimensions of sentiment within a sentence. However, the majority of existing sentiment analysis models predominantly focus solely on aspect extraction or sentiment tendency analysis, catering to single-task patterns. Addressing these issues, this paper introduces a fine-grained sentiment analysis model named BLAB (BERT Local Context Focus AD-BiReGU), which is founded on multi-task learning principles. In this model, the AD-BiReGU module is integrated into the BERT-LCF framework to enable simultaneous aspect word extraction and fine-grained sentiment analysis. Initially, the pre-trained BERT model captures initial features of both local and global contexts. Within the feature extraction layer, local context features are extracted through a local context focus mechanism combined with the multi-head attention mechanism, facilitating dynamic context feature masking. Simultaneously, the two-layer BiReGU model, grounded in the attention mechanism, incorporates context information into the neural network model, capturing long-term dependencies between labels and textual features to extract global features. Subsequently, local text information and global information are fused and input into the nonlinear layer to derive the ultimate sentiment polarity results. Comparative experiments indicate that integrating the AD-BiReGU module enhances performance noticeably for aspect word extraction tasks within multitasking scenarios.