Author(s) :

Moïse KASOMBO TSHIBANGU.

Volume/Issue :

Abstract :

Manual screening of thesis-topic redundancy and verification of similarities in academic work remain difficult at the Faculty of Sciences and Technologies of the University of Kinshasa because of the lack of tools adapted to the local context. This paper presents AcademiaCheck, a web platform combining two functions: intelligent validation of thesis topics and computer-assisted similarity detection in submitted work. The engine, implemented natively in TypeScript, combines TF-IDF vectorisation, cosine similarity, and the Jaccard index. This proof-of-concept study primarily evaluates topic validation using an anonymised corpus of 83 dossiers (57–110 words) compared with a real catalogue of 114 topics (snapshot at the evaluation date), including 27 documented redundancy cases across four difficulty levels. At the operational calibration (TF-IDF component weight = 0.70; threshold = 0.20), precision reaches 1.00 (95% CI: 0.86–1.00) and recall 0.89 (95% CI: 0.72–0.96); the F1-score is 0.94 (95% bootstrap CI: 0.86–1.00). A systematic grid search over 198 (weight, threshold) pairs places the retained calibration on the optimal plateau. Comparison with a multilingual semantic model (Sentence-BERT, paraphrase-multilingual-MiniLM-L12-v2) shows a slight recall gain (0.93), with approximately 50% higher computation time. The alternative-topic recommendation module is considered a functional prototype. Because the corpus is synthetic, these results characterise system performance under controlled conditions and support subsequent evaluation on real submissions.

No. of Downloads :

0