We assign scores to each answer choice by comparing the similarity between each reformed sentence and the corresponding article and background corpus, based on the features extracted from different annotators. These features could be defined by both low-level (e.g., bag of words) and structural (e.g., dependency parsing) information. We also need to come up with an appropriate model that combines the feature values and different types of scores to finally give the score of each candidate answer.
We assign scores to each answer choice by comparing the similarity between each reformed sentence and the corresponding article and background corpus, based on the features extracted from different annotators. These features could be defined by both low-level (e.g., bag of words) and structural (e.g., dependency parsing) information. We also need to come up with an appropriate model that combines the feature values and different types of scores to finally give the score of each candidate answer.