CRS-Mentor: A Conversational Recommender System for Large Language Model-Enhanced Academic Resource Mentorship
Zhichao Ma, Baisong Liu, Yunfei Li, Xueyuan Zhang · 2025
With the explosive growth of various academic information on the Internet, researchers' demand for academic resource recommendation systems that cater to personalized needs is becoming increasingly urgent. National Science Foundation report statistics From 2012 to 2022, the number of global scientific and engineering publications, including peer-reviewed journal articles and conference papers, increased at an average annual rate, with a total growth of 59% over the decade. This surge in publications has further intensified the challenge for researchers, who must dedicate considerable time and effort to navigate the expanding body of literature in their fields. However, faced with redundant search results, users are often overwhelmed and unable to find the information they need in the shortest possible time, which leads to traditional recommendation systems often failing to fully consider users' specific needs. To address this issue, we propose the CRS-Mentor model, which utilizes Large Language Models to engage users meaningfully, identify their needs, and provide personalized academic resource recommendations. The primary objective of designing this system is to aid students, researchers, and academic professionals in swiftly and precisely locating relevant literature amidst an extensive array of academic resources. These resources include journal articles, conference papers, academic works, and online databases. CRS-Mentor adopts a hierarchical intent understanding framework to distinguish academic queries from non-academic interactions, ensuring precise semantic parsing for complex research requests. This study evaluated CRS-Mentor using two different datasets, AMINER and DBLP, which compared its performance with traditional academic resource recommendation systems. The results indicate that CRS-Mentor provides more accurate and diverse recommendations by understanding the intent and context behind user queries. In addition, the system's ability to adjust recommendations based on user feedback and interactions over time enhances its effectiveness as a mentor tool for academic resource discovery.