Adapting BERT Embeddings for Text Correlation of Military Domain Specific Content

Abstract

This paper addresses the problem of how to achieve similarity search for text sequence correlation with a proper semantic foundation. Natural Language Processing (NLP) is fundamental for answering Community of Interest (COI) associated questions and this paper presents and compares three methods for similarity search. The methods are using Google introduced transformer models, BERT being the most well known. Combining techniques for pre-processing data, enhancing BERT and post-BERT adjustments are tested in an experimental setting and results are presented in this paper.