Detecting Novelty in Indonesian Research Grant Proposals Using Machine Learning and Natural Language Processing

Abstract

Assessing the novelty of research proposals represents a critical challenge in large-scale research grant management, particularly when the number of submitted proposals is substantial and topic diversity continues to increase. This study proposes an analytical approach to detect the novelty of Indonesian academic research proposal titles by leveraging semantic similarity based on Natural Language Processing (NLP). The CRISP-DM framework is adopted as the research methodology, with a comparative evaluation of three text representation approaches: TF-IDF, Latent Semantic Analysis (LSA), and IndoBERT. The dataset comprises 52,084 research proposal titles submitted between 2020 and 2024, which, after administrative deduplication, resulted in 44,429 unique titles. Inter-title similarity is computed using cosine similarity, followed by similarity distribution analysis and the application of a Top-K nearest neighbors strategy. In this study, novelty is conceptualized as novelty risk, defined as an analytical indicator of relatively low novelty derived from Top-1 similarity values and patterns of semantic similarity distribution. The results demonstrate that IndoBERT exhibits higher semantic sensitivity and a more stable similarity distribution compared to TF-IDF and LSA. Novelty risk estimation is conducted in an aggregated manner using rule-based analytical interpretation, without manual labeling or automated decision-making. The proposed approach functions as a decision support system to assist in the pre-screening of research proposals, while preserving the central role of expert reviewers in qualitative evaluation.