This will be a talk on ongoing work seeking to expand and improve
automatic short answer scoring in the language-learning context. We are
particularly interested in expanding current approaches to consider
semantic relationships between texts, questions, and answers. In this
talk we'll present both our work to date and a number of interesting
open research questions. Looking forward to your feedback!

Short answer questions are a common type of exercise in foreign language
learning, used for e.g. reading and listening comprehension tasks. In
automatic short answer grading, the goal is to learn a classification
that assigns labels (correct/incorrect, number of points etc.) to
student answers. These labels express whether the student answer is
semantically equivalent to a target answer, which is typically specified
by a teacher, regardless of the grammaticality of the learner answer.

We build on an existing short answer classification model by Meurers et
al. They use alignment on different linguistic levels (tokens, chunks,
dependency relations) to match a learner answer against a target answer,
but do not use the underlying texts as evidence for their
classification. We want to extend their classifier by not only matching
learner answers against target answers, but also employing e.g. textual
entailment methods to identify the text passages addressed by a learner
answer and use this additional information in the classification process.