Aspect based sentiment analysis with instruction tuning and external knowledge enhanced dependency graph

Aspect-Based Sentiment Analysis (ABSA) is generally defined as a fine-grained task in Natural Language Processing (NLP). Recently, the integration of the Large Language Model (LLM) and Graph Convolutional Network (GCN) has been widely studied to excavate the underlying contextual information and support the sentiment polarity prediction. However, in existing research, the LLM is usually employed directly to generate the contextual feature representation without any specific instructions, which is not suitable for learning the domain language corpus. In addition, the existing works usually fuse the contextual feature and graph feature by GCN simply, and it ignores further specific processing to highlight the sentiment representations before the model’s final outputting. To tackle these two imperfections, this work proposes a novel ABSA model Instruction Tuning-based Graph Convolutional Network (ITGCN) to implement the subtask of predicting sentiment polarities \(^\textrm{R2}\) ${}^{\text{R2}}$ , which leverages the instructed LLM to generate the task-oriented contextual representation and the GCN to exploit the external affective knowledge-assisted syntactic features. In the proposed ITGCN, firstly, the inputting sentence is reconstructed with the designed task-specific instructions, which tell the LLM what is the target in the input. Secondly, this work’s dependency graph, before being processed by GCN, is weighted by the affective knowledge extracted from SenticNet. This kind of dependency graph is endowed with affective information, which is closer to the intention of the related study. Finally, to learn more structured knowledge, a bi-layer sentiment representation module is proposed and utilized to enhance the feature representation. To validate the effectiveness of the proposed ITGCN, extensive experiments have been conducted on five public and available datasets. The proposed ITGCN achieves competitive performance and outperforms the selected state-of-the-art baselines, obviously.