Improved Local Citation Recommendation Based on Context Enhanced with Global Information

Abstract

Local citation recommendation aims at finding articles relevant for given citation context. While most previous approaches represent context using solely text surrounding the citation, we propose enhancing context representation with global information. Specifically, we include citing article's title and abstract into context representation. We evaluate our model on datasets with different citation context sizes and demonstrate improvements with globallyenhanced context representations when citation contexts are smaller.1 Most work on local CR uses contexts from published articles with citations masked out. While this is a somewhat artificial setup as those articles are biased toward existing citations, local CR can nonetheless be used in this setup to recommend additional citations.

Local Citation Recommendation with Hierarchical-Attention Text Encoder and SciBERT-based Reranking

“…Extensive experiments on the arXiv dataset as well as on previous datasets including ACL-200 [24], RefSeer [24,5], and FullTextPeerRead [15] show that our local citation recommendation system performs significantly better on both prefetching and reranking than the baseline and requires fewer prefetched candidates in the reranking step thanks to higher recall of our prefetching system, which indicates that our system strikes a better speed-accuracy balance.…”

“…The BERT-GCN model was evaluated on small datasets of only thousands of papers, partly due to the high cost of computating the GCN, which limited its scalability for recommending citations from large paper databases. Although recent studies [24,3,5,21] adopted the prefetching-reranking strategy to improve the scalability, the prefetch part (BM25 or TF-IDF) only served for creating datasets for training and evaluating the reranking model, since the target cited paper was added manually if it was not retrieved by the prefetch model to make the recall of the target among the candidate papers always equal to 1. Therefore, these recommendation systems were evaluated in an artificial situation with an ideal prefetching model that in reality does not exist.…”

“…Therefore, existing datasets on local citation recommendation are usually limited in size. For example, the ACL-200 [24] and the FullTextPeerRead [15] contain only thousands of papers. One of the largest datasets is RefSeer used in Medić and Šnajder [24], which contains 0.6 million papers in total, but this dataset is not up-to-date as it only contains papers prior to 2015.…”

“…In global citation recommendation, the query text includes the title and the abstract of a source paper [2]. In contrast, in local citation recommendation, the query consist of two sources of contexts [11,24]: 1) the text surrounding the citation placeholder with the information of the cited paper removed (the local context); and 2) the title and abstract of the citing paper as the global context. The aim of local citation recommendation is to find the missing paper cited at the placeholder of the local context.…”

“…The speed-accuracy tradeoff can be flexibly dealt with using a two-step prefetching-reranking strategy: 1) A fast prefetching model first retrieves a set of candidate papers from the database; 2) a more sophisticated model then performs a fine-grained analysis of scoring each candidate paper and reordering them to result in a ranked list of recommendations. In recent studies [24,3,5,21], (TF-IDF) [29] or BM25 [30] were generally used as the prefetching algorithm, which were neither fine-tuned nor taken into consideration when evaluating the recommendation performance.…”