A novel paper-reviewer recommendation method based on a semantics and correlation fusion model.
Peng Zhang,Peng Fu,Kaixuan Chen ,Jing Zhang
DOI: https://doi.org/10.1145/3661725.3661748
2024-01-01
Abstract:As academic research progresses, the blind review process of dissertations faces challenges in reviewer selection while ensuring independence, fairness, and objective evaluation. Existing systems and research often rely solely on text representation and vector similarity calculations. However, using text alone cannot accurately represent papers, and relying solely on similarity calculations between representations may not be comprehensive enough for matching.Therefore, we propose a reviewer recommendation method. Firstly, we introduce a representation method that integrates semantics and correlation (SCF) to represent the semantic information of reviewers and papers. This method uses large model label embedding to inject prior knowledge, enriching the semantic aspects of text embedding. Simultaneously, we construct two correlation graphs: one for paper information and one for reviewer majors. In the paper information correlation graph, we construct a graph of associations between dissertation subjects, research directions, and keywords. Using a graph attention network, we learn the associations between them, which are then fused with the semantic representation of paper titles to obtain paper embedding. In the reviewer majors correlation graph, we construct graphs for different subjects and obtain reviewer major embedding using the same method, thereby integrating semantic and correlation. Furthermore, we propose a matching method based on similarity matrices and clue vectors (SMC) to match papers and reviewers and subsequently select suitable reviewers. Firstly, considering the influence of different reviewer disciplinary backgrounds on matching results, we construct a similarity matrix based on representations of reviewer disciplinary information, representations of papers to be reviewed, and representations of papers related to reviewers. This matrix represents the complex matching states between papers to be reviewed and papers related to reviewers in different disciplinary backgrounds. Secondly, we extract fine-grained clue vectors from the similarity matrix to integrate complex matching states. Finally, we convert the paper reviewer recommendation problem into a classification problem based on clue vectors for classification prediction. Experiments on real dissertation submission data demonstrate that SCF-SMC outperforms existing methods in terms of performance.