Introduction
Local citation recommendation—the task of finding the most appropriate reference for a specific sentence or context—is a cornerstone of academic writing. Unlike global recommendation, which suggests papers based on an entire document's theme, local recommendation requires a nuanced understanding of specific claims. Current state-of-the-art models often fall short because they treat the query and candidate papers as isolated text blocks, neglecting the rich structural information inherent in citation networks and scientific taxonomies.
This paper introduces SymTax, a three-stage framework that mimics human research intuition. By leveraging "symbiotic" relationships within citation networks and fusing hierarchical taxonomies in hyperbolic space, SymTax achieves significant performance gains over existing methods.
The SymTax Framework
The SymTax pipeline is divided into three distinct phases: Prefetching, Enrichment, and Reranking.
- The Prefetcher: This initial stage acts as a high-speed filter. It uses a dual-encoder architecture to project queries and candidate papers into a shared vector space, performing a first-pass retrieval based on cosine similarity.
- The Enricher (Symbiosis Module): Inspired by biological symbiosis, this module expands the candidate pool by exploring the "ego network" of the top-ranked papers. It identifies "mutualistic" relationships (where both a paper and its references are relevant) and "parasitic" ones (where a reference is more relevant than the paper citing it), effectively capturing how researchers naturally discover new work.
- The Reranker (Taxonomy Fusion): The final stage performs a deep, fine-grained comparison. It utilizes a Fusion Multiplexer to integrate flat categories (e.g., arXiv tags) with hierarchical structures (e.g., the ACM Computing Classification System). By projecting these taxonomies into hyperbolic space, the model preserves the exponential growth of hierarchical trees, leading to more accurate semantic representations.
The ArSyTa Dataset
To evaluate SymTax at scale, the authors introduced ArSyTa, a massive new benchmark for citation recommendation. Containing 8.27 million citation contexts across nearly half a million papers, ArSyTa is significantly larger and denser than previous datasets like RefSeer or ACL-200. It includes rich metadata, such as publication years and section headings, providing a more challenging and realistic environment for testing recommendation systems.
Experimental Results
SymTax was evaluated against several state-of-the-art models, including BM25, SciNCL, and HAtten, across five diverse datasets.
As shown in the results table, the SymTax (SciV) variant consistently outperforms all baselines. Notable achievements include:
- Recall@5 Improvements: Gains of 26.66% on ACL-200 and 39.25% on RefSeer compared to the previous best-performing models.
- ArSyTa Benchmark: On the new massive dataset, SymTax achieved a 22.56% improvement in Recall@5, demonstrating its robustness on large-scale, modern scientific corpora.
Discussion and Key Insights
The success of SymTax highlights several critical factors in citation recommendation. First, the Enricher module proved that citation networks are a powerful signal; simply including a candidate's references improved Recall@5 by up to 39%. Second, the use of hyperbolic geometry was essential for representing complex taxonomies without the distortion common in Euclidean space. Finally, the choice of language model matters; SciBERT, pre-trained on scientific literature, generally provided better results than general-purpose models when integrated into the SymTax architecture.
Conclusion
SymTax represents a significant step forward in mimicking human-like discovery in citation recommendation. By combining text-based retrieval with structural network signals and hierarchical knowledge, it provides a more effective tool for researchers. Future work will focus on optimizing the Reranker for real-time online applications and expanding the taxonomy fusion to include non-computer science domains.
Summary updated