An Approach for Evaluating Semantic Similarity in Research Papers via Siamese BERT Architecture

EasyChair Preprint 15448

Abstract

Document similarity analysis is critical for various NLP tasks like information retrieval and plagiarism detection. Traditional methods based on word-to-word mapping struggle with capturing contextual nuances. Existing solutions lack the capability to provide domain-specific accuracy and enriched search experiences. One such field is finding similar research papers. Often researchers struggle to find papers similar to a certain paper and have to rely on basic keyword-based search. This hinders to provide the best match based on the overall context. In this work, we propose a novel methodology that integrates BERT with a Siamese Neural Network to capture semantic textual similarity of research papers. Our approach goes beyond simple similarity evaluation by conducting a nuanced semantic search of overall context and provides a representative similarity score. This offers a more accurate and refined search experience. Furthermore, we curate a dataset of over 10,000 NLP research paper abstracts to train our model. The model excels in identifying the contextual relationships between documents, making it highly effective for domain-specific applications. This model can significantly improve the user experience in document retrieval systems, particularly for academic research and recommendation.

Keyphrases: BERT, Data Science, NLP, Siamese Neural Network, semantic similarity

@booklet{EasyChair:15448,
  author    = {Pritam Sarkar and Soumyaneel Sarkar and Wazib Ansar and Amlan Chakrabarti},
  title     = {An Approach for Evaluating Semantic Similarity in Research Papers via Siamese BERT Architecture},
  howpublished = {EasyChair Preprint 15448},
  year      = {EasyChair, 2024}}