Academic Aggregated Search Approach Based on BERT Language Model

Abstract

Academic Search concerns the indexing and retrieval of information objects like tutorials, blog posts, papers, books, etc. in the domain of academic research. When looking for a specific information, many researchers can't find all the information needed on the topic for which they are searching in the same academic search engine, which pushes them to search in each one separately. Therefore, we propose to exploit the task of aggregating search results from various sources in an academic domain. Exploiting the advantage of such systems may improve retrieval performance of academic search results by selecting the most appropriate vertical results according to the user's information need, and presenting them in a single page. Our proposed approach consists of exploiting BERT contextual embeddings in unsupervised fashion and taking advantage of their rich semantic features to improve the quality of the retrieved results. To the best of our knowledge, this is the first time when aggregated search paradigm is investigated in an in academic environment using BERT embeddings. Experimental results show that by only using BERT algorithm as a core component in our system, has given excellent results for our task and helped us achieved our objective in this work. In addition, a comparison with other state-of-the-art language models such as ELMo, USE and XLNet was made. The findings showed that the best performance was obtained using the BERT language model, which proved that using BERT model in our academic aggregated search system is a good choice.