Analyzing short-answer questions and their automatic scoring - studies on semantic relations in reading comprehension and the reduction of human annotation effort

Andrea Horbach

Abstract: Short-answer questions are a wide-spread exercise type in many educational areas. Answers given by learners to such questions are scored by teachers based on their content alone ignoring their linguistic correctness as far as possible. They typically have a length of up to a few sentences. Manual scoring is a time-consuming task, so that automatic scoring of short-answer questions using natural language processing techniques has become an important task. This thesis focuses on two aspects of short-answer questions and their scoring: First, we concentrate on a reading comprehension scenario for learners of German as a foreign language, where students answer questions about a reading text. Within this scenario, we examine the multiple relations between reading texts, learner answers and teacher-specified target answers. Second, we investigate how to reduce human scoring workload by both fully automatic and computer-assisted scoring. The latter is a scenario where scoring is not done entirely automatically, but where a teacher receives scoring support, for example, by means of clustering similar answers together. Addressing the first aspect, we conduct a series of corpus annotation studies which highlight the relations between pairs of learner answers and target answers, as well as between both types of answers and the reading text they refer to. We annotate sentences from the reading text that were potentially used by learners or teachers for constructing answers and observe that, unsurprisingly, most correct answers can easily be linked to the text; incorrect answers often link to the text as well, but are often backed up by a part of the text not relevant to answer the question. Based on these findings, we create a new baseline scoring model which considers for correctness whether learners looked for an answer in the right place or not. After identifying those links into the text, we label the relation between learner answers and target answers as well as between reading texts and answers by annotating entailment relations. In contrast to the widespread assumption that scoring can be fully mapped to the task of recognizing textual entailment, we find the two tasks to be only closely related and not completely equivalent. Correct answers do often, but not always, entail the target answer, as well as part of the related text, and incorrect answers do most of the time not stand in an entailment relation to the target answer, but often have some overlap with the text. This close relatedness allows us to use gold-standard entailment information to improve the performance of automatic scoring. We also use links between learner answers and both reading texts and target answers in a statistical alignment-based scoring approach using methods from machine translation and reach a performance comparable to an existing knowledge-based alignment approach. Our investigations into how human scoring effort can be reduced when learner answers are manually scored by teachers are based on two methods: active learning and clustering. In the active learning approach, we score particularly informative items first, i.e., items from which a classifier can learn most, identifying them using uncertainty-based sample selection. In this way, we reach a higher performance with a given number of annotation steps compared to randomly selected answers. In the second research strand, we use clustering methods to group similar answers together, such that groups of answers can be scored in one scoring step. In doing so, the number of necessary labeling steps can be substantially reduced. When comparing clustering-based scoring to classical supervised machine learning setups, where the human annotations are used to train a classifier, supervised machine learning is still in the lead in terms of performance, whereas clusters provide the advantage of structured output. However, we are able to close part of the performance gap by means of supervised feature selection and semi-supervised clustering. In an additional study, we investigate the automatic processing of learner language with respect to the performance of part-of-speech (POS) tagging tools. We manually annotate a German reading comprehension corpus both with spelling normalization and POS information and find that the performance of automatic POS tagging can be improved by spell-checking the data using the reading text as additional evidence for lexical material intended in a learner answer.