A Framework for Evaluation of Machine Reading Comprehension Gold Standards

Machine Reading Comprehension (MRC) is the task of answering a question over paragraph of text. While neural MRC systems gain popularity and achieve performance, issues are being raised with the methodology used to their performance, particularly concerning the data design of gold that are used to evaluate them. There is but a limited understanding the challenges present in this data, which makes it hard to draw comparisons formulate reliable hypotheses. As a first step towards alleviating the, this paper proposes a unifying framework to systematically investigate present linguistic features, required reasoning and background knowledge factual correctness on one hand, and the presence of lexical cues as a bound for the requirement of understanding on the other hand. We propose qualitative annotation schema for the first and a set of approximative for the latter. In a first application of the framework, we analyse MRC gold standards and present our findings: the absence of features contribute towards lexical ambiguity, the varying factual correctness of expected answers and the presence of lexical cues, all of which potentially the reading comprehension complexity and quality of the evaluation data.