BERT-Based Semantic Retrieval for Academic Abstracts

This research focuses on improving the academic abstract search system to help reduce search time and facilitate users in searching with general terms, without the need for specific terminology, and using synonyms in the search. We evaluated the performance of four techniques: TF-IDF, Doc2Vec, Word2Vec, and BERT, through experiments using a dataset of 10,000 abstracts in the field of science and technology, which were embedded using the four techniques. The main objective is the accuracy of using keywords to search for the four techniques, which then yield abstracts that correctly match the 100 prepared abstracts before the search. This serves as a basis for developing an efficient techniques for future academic abstract search systems.