Similarity Detection of Natural-Language Questions and Answers using the VANiLLa dataset

Abstract

Abstract Question Answering refers to the task of providing a matching answer for a given user’s natural-language question. We assume here that the received response from a Question Answering system is also given as a natural-language output. From this scenario the task is raised to validate if the received natural-language answer is valid w.r.t. the given question. In this paper, we will present our approach to compute the similarity of a question to the corresponding natural-language answer while using the features of the surface form of the question and the answer. Our metrics is the distinction between a matching answer and inappropriate answers to a given question. Hence, our research agenda is dedicated to improving the answer quality of Question Answering systems by using the natural-language representation of its input and output.