Sentence extraction for machine comprehension

Abstract

Machine comprehension is a broad research area from Natural Language Processing domain, which deals with making a computerised system understand the given natural language text. Question answering system is one such variant used to find the correct ‘answer’ for a ‘query’ using the supplied ‘context’. Using a sentence instead of the whole context paragraph to determine the ‘answer’ is quite useful in terms of computation as well as accuracy. Sentence selection can, therefore, be considered as a first step to get the answer. This work devises a method for sentence selection that uses cosine similarity and common word count between each sentence of context and question. This removes the extensive training overhead associated with other available approaches, while still giving comparable results. The SQuAD dataset is used for accuracy based performance comparison.