Scientific Paper Recommendation System: Application of Sentence Transformers and Cosine Similarity Using arXiv Data

Searching for relevant scientific literature faces complex challenges due to the proliferation of academic publications. This research develops a semantic similarity-based scientific paper recommendation system by utilizing Sentence Transformer (all-MiniLM-L6-v2 model) and cosine similarity algorithm on arXiv dataset (15,504 papers in Computer Science). The system is implemented as a Streamlit-based interactive web application that accepts user queries and recommends related papers based on semantic similarity. Performance evaluation using Precision, Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG) metrics showed that embedding text from the Introduction section without pre-processing yielded the best performance (NDCG: 0.7590; MAP: 0.6960; MRR: 0.7254), outperforming Abstract-based or text combination approaches. A user test of 45 respondents confirmed the effectiveness of the system: 95.5% expressed satisfaction with the relevance of the recommendations, and 93.3% confirmed a significant reduction in manual search time. The findings prove that retaining the raw text structure in the Introduction is optimal for semantic representation. Development suggestions include multidomain dataset expansion and transformer model optimization for accuracy improvement.