ABSTRACT Recommender systems play a crucial role in supporting scholarly research by helping researchers navigate the rapidly growing volume of academic literature. However, most of the existing academic recommender systems suffer from limitations including over‐dependency on keyword matching, weak contextual understanding, and a lack of contextual relevance that significantly reduce the accuracy and usefulness of the retrieved results. To address these challenges, the SemanticRec framework is proposed that utilizes traditional similarity measures with contextual semantic embeddings based on a domain‐specific language model. The proposed approach uses cosine similarity and Jaccard similarity as the baseline lexical search methods and also utilizes SciBERT embeddings to improve the contextual relevance during document retrieval. SemanticRec is evaluated on large academic datasets (1.7M DBLP, 1M ACM records) to assess robustness and generalizability across multiple metrics. Experimental results show that the precision of cosine similarity with lemmatization enhances from 40.5% to 48.0%, while the F1‐score improves from 37.5% to 55.0%. Likewise, the F1‐score of the Jaccard similarity increases from 27.25% to 55.5%. This finding indicates that the proposed SemanticRec system improves contextual relevance as well as the efficiency of retrieval in the academic literature discovery.
Ahmad et al. (Wed,) studied this question.