A novel automated framework for fine-grained sentiment analysis of application reviews using deep neural networks

The substantial volume of user feedback contained in application reviews significantly contributes to the development of human-centred software requirement engineering. The abundance of unstructured text data necessitates an automated analytical framework for decision-making. Language models can automatically extract fine-grained aspect-based sentiment information from application reviews. Existing approaches are constructed based on the general domain corpus, and are challenging to elucidate the internal technique of the recognition process, along with the factors contributing to the analysis results. To fully utilize software engineering domain-specific knowledge and accurately identify aspect-sentiment pairs from application reviews, we design a dependency-enhanced heterogeneous graph neural networks architecture based on the dual-level attention mechanism. The heterogeneous information network with knowledge resources from the software engineering field is embedded into graph convolutional networks to consider the attribute characteristics of different node types. The relationship between aspect terms and sentiment terms in application reviews is determined by adjusting the dual-level attention mechanism. Semantic dependency enhancement is introduced to comprehensively model contextual relationships and analyze sentence structure, thereby distinguishing important contextual information. To our knowledge, this marks initial efforts to leverage software engineering domain knowledge resources to deep neural networks to address fine-grained sentiment analysis issues. The experimental results on multiple public benchmark datasets indicate the effectiveness of the proposed automated framework in aspect-based sentiment analysis tasks for application reviews.