Semantic Search with BERT

In this workflow, abstracts from the COVID-19 Open Research Dataset ( are read in to perform semantic search using a TensorFlow 2 BERT model. For this purpose, a BERT model that has already been trained on the CORD-19 dataset is loaded from TensorFlow Hub ( The BERT embeddings created from the abstracts are used to find semantically similar abstracts for the question asked; they are used to calculate the cosine similarity to the query embeddings and the semantically most relevant papers are displayed in a view afterwards. The data can be downloaded from

This is a companion discussion topic for the original entry at