Search

When writing academic papers, referencing statements, claims and previous studies is always an important activity. However, it is considered challenging for scientists to find relevant and appropriate scientific articles which are closely related to their current works in order to reference them in their research. As a consequence of the rapid growth in scientific papers being published every year, researchers might easily get overwhelmed within a huge number of resources. One way to help them find the desired references more easily is to use context-aware citation recommendation. The citation recommender system can automatically provide a list of suitable papers as references based on specified inputs which reflect the researchers' interests. Among the outstanding achievements of deep learning and natural language processing in recent years, the utilization of deep neural learning architectures have supported to address the problem of citation recommendation. As the result, the neural citation recommendation area has received much attention from the academic community, with the aim of enhancing the precision and correctness of the results of existing citation recommendation systems. Following this research direction, in our paper we present a novel context-aware citation recommendation model, called RHN-DualLCR (Recurrent Highway Networks – Dual Local Citation Recommendation), which integrates Recurrent Highway Networks (RHN), an improved model of the original Bidirectional Long Short-Term Memory (BiLSTM) architecture, and uses a SciBERT-based (Science text of Bidirectional Encoder Representations from Transformers) embedding layer to build up the efficiency of the state-of-the-art local citation recommendation model, which enriches context representation with global information. Our research demonstrates its originality and relevance because we used one of the latest achievements of deep models (RHN model) and natural language processing (SciBERT) applied to the citation recommendation problem. We have conducted experiments on the RHN-DualLCR model on 3 widely known datasets for the citation recommendation problem: ACL-200 (Association for Computational Linguistics), ACL-600 and RefSeer, and also used 2 common evaluation standards Mean Reciprocal Rank (MRR) and the Recall@K (R@K for short) to evaluate the performance of our model. Experimental results show that our proposed model is 3% to 16% better than the original models or state-of-the-art models. [ABSTRACT FROM AUTHOR]

In the era of rapid paper publications in various venues, automatic citation recommendations would be highly useful to researchers when they write papers. Local citation recommendation aims to recommend possible papers to cite given local citation contexts. Previous work mainly computes the similarity score between citation contexts and cited papers on a one-to-one basis, which is quite time-consuming. We train a pair of neural network encoders that map citation contexts and all possible cited papers to the same vector space, respectively. After that, we index the positions of all cited papers in the vector space. This makes our process for searching recommended papers considerably faster. On the other hand, existing methods tend to recommend papers that are highly similar to each other, which makes recommendations lack diversity. Therefore, we extend our algorithm to perform multi-topic recommendations. We generate multi-topic training examples based on the index we mentioned earlier. Furthermore, we specially design a multi-group contrastive learning method to train our model so that it can distinguish different topics. Empirical experiments show that our model outperforms previous methods by a wide margin. Our model is also light weighted and has been deployed online so that researchers can use it to obtain recommended citations for their own paper in real-time. • Proposed FLCR algorithm with sentence-transformer & k-d tree. • Introduced multi-topic citation recommendation for diverse contexts. • Developed large-scale dataset with 1.7 million citation contexts for evaluation. • Demonstrated significant performance improvement over previous methods. • Deployed demo for real-time citation recommendations. [ABSTRACT FROM AUTHOR]

As the number of published research articles grows on a daily basis, it is becoming increasingly difficult for scientists to keep up with the published work. Local citation recommendation (LCR) systems, which produce a list of relevant articles to be cited in a given text passage, could help alleviate the burden on scientists and facilitate research. While research on LCR is gaining popularity, building such systems involves a number of important design choices that are often overlooked. We present an empirical study of the impact of the three design choices in two-stage LCR systems consisting of a prefiltering and a reranking phase. In particular, we investigate (1) the impact of the prefiltering models' parameters on the model's performance, as well as the impact of (2) the training regime and (3) negative sampling strategy on the performance of the reranking model. We evaluate various combinations of these parameters on two datasets commonly used for LCR and demonstrate that specific combinations improve the model's performance over the widely used standard approaches. Specifically, we demonstrate that (1) optimizing prefiltering models' parameters improves R @ 1000 in the range of 3% to 12% in absolute value, (2) using the strict training regime improves both R @ 10 and M R R (up to a maximum of 3.4% and 2.6%, respectively) in all combinations of dataset and prefiltering model, and (3) a careful choice of negative examples can further improve both R @ 10 and M R R (up to a maximum of 11.9% and 8%, respectively) in both datasets used Our results show that the design choices we considered are important and should be given greater consideration when building LCR systems. • Certain design choices affect the accuracy of local citation recommendation systems. • Training reranker model with a strict regime improves the model's performance. • Triplet-based reranking models benefit from non-random negative sampling strategies. • The best negative sampling strategy for triplet construction depends on the dataset. [ABSTRACT FROM AUTHOR]

• A deep-learning-based citation recommendation model is proposed. • The proposed model is called dual encoder retrieval (DER) model. • The DER model can deeply understand the semantics of sentences. • The citation recommendation of DER employs sequence query, instead of keywords. Citation recommendation recommends relevant documents to users based on their inputs and other information. Many traditional citation recommendation models use keywords to describe item attributes and ignore the semantics of sequences, which cause the relevance of the search results unsatisfactory. This paper proposes a deep-learning-based dual encoder retrieval (DER) model, which combines a text representation technique and a sentence pair matching approach, to improve the performance of citation recommendation. First, an input query and paper titles from publication databases are encoded to semantic vectors separately by two deep-learning-based encoders. Second, the semantic vector of the input query is matched with vectors that representing papers in the published databases by the multilayer perceptron approach to compute similarity scores. Finally, a list of documents, which are sorted in descending order of similarity scores, is generated. To validate the effectiveness of the proposed approach, it is compared with five baselines using a citation dataset. The results show that the proposed model achieves the best performance in terms of accuracy, recall, F1-measure, and AUC. In addition, we compare the DER (Glove) model with Google Scholar using a small example of twenty articles. The DER (Glove) model outperformed Google Scholar in seven recommendations, and tied in ten recommendations. [ABSTRACT FROM AUTHOR]

• Recommend citations from the perspective of cited paper instead of citing paper. • Design four forms of citation content for different citation motivations. • Design 132 methods for generating semantic representation of cited paper. • Find the most suitable form and method through quantitative comparisons. Citation recommendation can help researchers quickly find supplementary or alternative references in massive academic resources. Current research on citation recommendation mainly focuses on the citing papers, resulting in the enormous cited papers are ignored, including the relations among cited papers and their citation context cited in citing papers. Moreover, cited paper's content is often denoted with its original title and abstract, which is hard to acquire and rarely considers different citation motivations. Furthermore, the most appropriate method for semantic representation of cited papers' relations and content is uncertain. Therefore, this paper studies citation recommendation from the perspective of semantic representation of cited papers' relations and content. Firstly, four forms of citation context are designed and extracted as cited papers' content considering citation motivations, as well as co-citation relationships are extracted as cited papers' relations. Secondly, 132 methods are designed for generating semantic vector of cited paper, including four network embedding methods, 16 methods by combining four text representation algorithms with four forms of citation content, and 112 fusion methods. Finally, similarity among cited papers is calculated for citation recommendation and a quantitative evaluation method based on link prediction is designed, to find the most appropriate form of citation content and the optimal method. The result shows that doc2vecC (Document to Vector through Corruption) with the form of CS&SS (Current Sentences and Surrounding Sentences) performs best, in which the AUC (Area Under Curve) and MAP (Macro Average Precision) reach 0.877 and 0.889 and have increased by 0.462 and 0.370 compared with the worst-performing method. This performance is slightly improved by parameters adjustment, and a case study is performed whose results have further proved the effectiveness of this method. In addition, among four forms of cited papers' content, CS&SS performs best in almost all methods. Furthermore, the fusion methods not always perform better than the single methods, where doc2vecC (CS&SS) performs better than the best fusion method GCN (Graph Convolutional Network). These results not only prove the effectiveness of citation recommendation from the perspective of cited paper, but also provide helpful and useful suggestions for method selection and citation content selection. The data and conclusions can be extended to other text mining-related tasks. Simultaneously, it is a preliminary research which needs to be further studied in other domains using emerging semantic representation methods. [ABSTRACT FROM AUTHOR]

• We construct a Heterogeneous Bibliographic Network, which exploits semantics among the network objects. • We propose a GAN-based network embedding model to address the network sparsity problem. • We propose a citation recommendation model to produce personalized results corresponding to researchers' preferences. • We conduct extensive experiments on two real-world datasets and prove the significance of the proposed model. The variety and plethora of research papers available on the Web motivated researchers to propose models that could assist users with personalized citation recommendations. In recent years, citation recommendation models using Network Representation Learning (NRL) methods have shown promising results. Nevertheless, existing NRL-based models are limited in terms of exploiting semantic relations and contextual information between the objects of bibliographic papers' networks. Additionally, these models cannot adequately explore the structure of heterogeneous information networks, topical relevance, and relevant semantics. Consequently, they suffer from network sparsity and inadequate personalization problems. To overcome these shortcomings, we present a network embedding model termed as Global Citation Recommendation employing Generative Adversarial Network (GCR-GAN). The proposed model exploits the Heterogeneous Bibliographic Network (HBN) to generate personalized citation recommendations. In particular, the proposed model utilizes semantic relations corresponding to the objects of the heterogeneous bibliographic network and captures network structure proximity employing the Scientific Paper Embeddings using Citation-informed Transformers (SPECTER) and Denoising Auto-encoder networks to learn semantic-preserving graph representations. Compared to baseline models, the recommendations generated by our model over the DBLP and ACM datasets prove that it outperforms baseline methods by gaining almost 11% and 12% improvement in terms of Mean Average Precision (MAP) and Normalized Discounted Cumulative Gain (nDCG) metrics, respectively. Furthermore, we analyzed the effectiveness of the proposed model considering network sparsity issue, where our model gains almost 7% better recall@100 score against the second-best counterpart. [ABSTRACT FROM AUTHOR]

The huge amount of research papers on the web makes finding a relevant manuscript a difficult task. In recent years many models were introduced to support researchers by providing personalized citation recommendations. Moreover, deep learning methods have been employed in this domain to improve the quality of the final recommendations. However, a thorough study that classifies citation recommendation models and examines their (a) strengths and weaknesses, (b) evaluation metrics used, (c) popular datasets, and challenges faced is missing. Therefore, with this survey, we present a new classification approach for deep learning models that provide citation recommendation. Our approach uses the following six criteria: data factors, data representation methods, methodologies, types of recommendations used, problems addressed, and personalization. Additionally, we present a comparative analysis of those models that use the same set of evaluation metrics and datasets. Moreover, we examine hot upcoming issues and solutions in light of explored literature. Also, the survey discusses and analyzes the evaluation metrics and datasets adopted by the explored models. Finally, we conclude our survey with trends and future directions to further assist research on that domain. • We propose a novel taxonomy for deep learning-based citation recommendation models. • We examine and analyze the most popular evaluation metrics and datasets. • We present a comparative analysis among the explored models. • We highlight prominent research issues and recommendations. • We uncover research trends and present future research directions. [ABSTRACT FROM AUTHOR]

Researchers face millions of research papers on various digital libraries. Therefore, finding relevant research work that meets the preferences of a researcher is a challenging task. Hence, different paper recommendation models have been proposed to address this issue. However, these models lack in exploiting prominent information factors, namely: papers' citations proximity, authors' information, papers' topical relevance, venues' information, researchers' preference dynamics, and labels information to produce quality recommendations. Additionally, these models encounter problems such as cold start papers and data sparsity. To overcome these problems, this paper presents a weighted probabilistic paper recommendation model termed as PR-HNE, which jointly learns researchers' and papers' dynamics by encoding information from six information networks into a joint latent space. Specifically, it captures papers' citation proximity, authors' collaboration proximity, venues' information, labeled information, and topical relevance to generate personalized paper recommendations. Compared to state-of-the-art models, the results generated by PR-HNE over publicly available datasets prove 4% and 6% improvement in Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR) metrics, respectively. Further, in the cold-start papers problem, the proposed model produced 8% better recall score than its counterparts. [ABSTRACT FROM AUTHOR]