A Sentence Similarity Approach for Machine Reading Comprehension System

Abstract

The growth of digital textual information demands Machine Reading (MR) research. The outcome of MR is text understanding. The level of understanding or its correctness is the biggest challenge in MR research. The Machine Reading Comprehension System (MRCS) with Multiple Choice Questions (MCQ) is considered to be one of the best applications for MR evaluation. The given comprehension is represented as Linguistic Knowledge Document (LKD) and the MCQ are generated as Candidate Answer Sentences (CAS) based on heuristic rules in our earlier works. Here we proposed a sentence similarity approach for finding the correct answer option for the MCQ. In our work we used Jaccard similarity measure for finding the similarity score. The similarity score between MCQ and each sentence of comprehension text is calculated. Each sentence has set of sentences in LKD. The similarity score between each generated sentence in LKD and MCQ are calculated. The highest similarity score obtained is selected as similarity score of the sentence in comprehension text. In the same way similarity score is calculated for each and every sentences in comprehension text. The sentence which obtained highest score is found to have highest relevance. Next the similarity score for CAS and LKD is done in similar way and the results are stored in a two dimensional matrix where row represent CAS and column represent sentences of comprehension text. Finally, we found the correct answer by our proposed decision making algorithm. The system is implemented using RACE dataset and it achieved good results.