Abstract

Make machines to read and comprehend information from natural language documents are not an easy task. Machine reading comprehension is a solution to alleviate this issue by extracting the relevant information from the corpus by posing a question based on the context. The problem associated with this knowledge retrieval is in the correct answer extraction from the context with language understanding. The traditional rule-based, keyword search and deep learning approaches are inadequate to infer the right answer from the input context. The Transformer based methodologies are used to excerpt the most accurate answer from the context document. This article utilizes one of the exceptional transformer models - BERT (Bidirectional Encoder Representations from Transformers) for empirical analysis for Neural Machine Reading Comprehension. This article aims to reveal the differences between the BERT and the domain-specific models. Furthermore, explores the need for domain specific models and how these models outperform the BERT.