Enhancing the hatten model for local citation recommendation using BiLSTM and attention pooling

Abstract

Over the past decade, citation recommendation has gained increasing attention due to the exponential growth of scientific publications. Among various approaches, local citation recommendation (LCR) - a content-based method leveraging textual context - has proven effective but faces scalability challenges when applied to large databases. To balance computational efficiency and accuracy, recent systems adopt a two-stage pipeline: a lightweight prefetching phase followed by a refined reranking stage. Building upon this direction, This paper proposes Enhanced-HAtten, an improved version of the HAtten-SciBERT model [1]. The proposed model retains the original two-stage architecture but augments the prefetching phase with a Bidirectional Long Short-Term Memory (BiLSTM) layer and attention pooling, enabling richer sequential and semantic representations. Experiments on two benchmark datasets-ACL-200 and FullTextPeerRead demonstrate that Enhanced-HAtten consistently outperforms the original HAtten-SciBERT pipeline, yielding over 10% improvement in both Mean Reciprocal Rank (MRR) and Recall@K, confirming its effectiveness for large-scale scholarly recommendation tasks.