Transformer-aware sequence-to-sequence network for personalized tag recommendation in software information sites

Abstract

Context: Automatic tag recommendation is crucial for content understanding and retrieval on software information sites. Existing approaches formulate tag recommendation as either multi-label classification problem or sentence matching. However, multi-label classification by treating tags as independent labels, neglects semantic relationships and dependencies among them, leading to inconsistent recommendations. Sentence matching techniques, which rely on lexical similarity, fail to capture contextual information and broader semantic meaning. Moreover, several works leverage limited content sources such as title, body, and code snippet of software objects, leading to data sparsity issues. Extant research provides generic tag suggestions, overlooking users’ expertise and interests. Objective: To address these limitations, we propose a novel trnsformer-based sequence-to-sequence framework for personlaised tag recommendation, dubbed as ANNOTATION. Methods: This approach enables the model to learn dependencies between tags and generate contextually relevant recommendations. To enhance the representation of software objects and mitigate data sparsity, we incorporate valuable information from associated comments, such as clarifications, usage examples, and bug reports. This additional conversational context provides insights into the problem, alternative solutions, and related concepts discussed by the community, resulting in more informed tag recommendations. Furthermore, we personalize tag suggestions by incorporating user profile descriptions and badges, which reflect users’ expertise and interests within specific domains. This ensures that generated tags align with both the software object and the user’s specific knowledge domain, contributing to a more tailored user experience. Results: Extensive empirical and qualitative evaluations on datasets from Code Review and Stack Overflow demonstrate that our approach significantly outperforms state-of-the-art methods. Conclusion: Our findings highlight the importance of considering tag dependencies, contextual information, and user preferences for accurate and personalized tag recommendation in software information sites. © 2026 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.