Automatic short-answer grading has been studied for more than a decade. The technique has been used for implementing auto assessment as well as building the assessor module for intelligent tutoring systems. Many early works automatically grade mainly based on the similarity between a student answer and the reference answer to the question. This method performs well for closed-ended questions that have single or very limited numbers of correct answers. However, some short-answer questions ask students to express their own thoughts based on various facts; hence, they have no reference answers. Such questions are called semi-open-ended short-answer questions. Questions of this type often appear in reading comprehension assessments. In this paper, we developed an automatic semi-open-ended short-answer grading model that integrates both domain-general and domain-specific information. The model also utilizes a long-short-term-memory recurrent neural network to learn the representation in the classifier so that word sequence information is considered. In experiments on 7 reading comprehension questions and over 16,000 short-answer samples, our proposed automatic grading model demonstrates its advantage over existing models.

This paper introduces a system that supports student-centered online one-to-one tutoring and evaluates the practical value of the system by running an experiment with 64 experienced mathematics teachers and 810 students in Grade 7. The experiment lasted for 50 days. A comprehensive evaluation was performed using students’ academic performance before and after usage of the system and the system log files. By classifying the students into active and inactive usage groups, it was determined that active students significantly outperformed inactive students on posttests, but with a small effect size. The results also suggested that high prior knowledge students tended to benefit more from using the system than low prior knowledge students. An explanation for this result was that students with a high level of prior knowledge were more likely to have good-quality interactions with their teachers. Therefore, although some advantages of this type of student-centered online one-to-one tutoring are observed, in this system, both the students and the teachers need to be further facilitated to produce more effective tutoring interactions.

Despite their drawback, multiple-choice questions are an enduring feature in instruction because they can be answered more rapidly than open response questions and they are easily scored. However, it can be difficult to generate good incorrect choices (called “distractors”). We designed an algorithm to generate distractors from a semantic network for four types of multiple choice questions in biology. By recruiting 200 participants from Amazon Mechanical Turk, the machine-generated distractors were compared to human-generated distractors in terms of question difficulty, question discrimination and distractor usefulness. The machine-generated and human-generated distractors performed very closely on all the three measures, suggesting that generating distractors from a semantic network for simple multiple choice questions is a viable method.

This project aimed to improve students’ learning and task performance using a non-cognitive learning companion in the context of both a tutor and a meta-tutor. The tutor taught students how to construct models of dynamic systems and the meta-tutor taught students a learning strategy. The non-cognitive learning companion was designed to increase students’ effort and persistence in using the learning strategy. It decided when to intervene and what to say using both log data and affective state monitoring via a facial expression camera and a posture sensor. Experiments with high school students showed that the non-cognitive learning companion increased students’ learning and performance. However, it had no effect on performance during a transfer phase in which the learning companion, meta-tutor, and tutor were all absent. The transfer phase null effect must be interpreted with caution due to low power, a possible floor effect, and other issues.