CITAR: Citation Inference using Taxonomy-Aligned Retrieval and Context-Aware Entity Matching

In the rapidly expanding landscape of academic research, efficiently retrieving relevant scholarly papers is a significant challenge. Modern existing citation recommendation systems and algorithms fall short on two practical dimensions: they treat hierarchical domain taxonomies as flat metadata rather than first-class retrieval structures, and they return whole-document suggestions without identifying where in a draft a citation should be inserted. To address these gaps, we present CITAR, a taxonomy-aligned, context-aware citation recommendation algorithm that fuses multi-taxonomy Directed Acyclic Graph (DAG) representations with hierarchical entity matching and semantic retrieval. For each input text segment, CITAR constructs a query DAG aligned across arXiv and ACM-CCS taxonomies, performs top-down entity matching against precomputed document DAGs, and returns both ranked citations and precise insertion offsets. We describe our corpus augmentation and held-out validation procedure for threshold selection, present an ablation study quantifying each module's contribution, and compare CITAR against BM25, SciNCL, and HAtten with full runtime and implementation details. We further demonstrate a lightweight post-hoc MMR re-ranker that enforces corpus-proportional exposure quotas and report exposure-disparity metrics to show how retrieval relevance and equitable visibility can be balanced in practice. On a 10,000-passage test set drawn from arXiv 2024 and 2025 papers, CITAR achieves Recall@10\,=\,0.43, MRR\,=\,0.25, and NDCG@10\,=\,0.3 at a mean query latency of 180\,ms.