To study the interaction between learners’ affective states, cognitive strategies, and learning gain in OELEs, I am collecting data from facial emotion detectors, eye gaze patterns, log data, videos from screen capture software, and human observation of affective states to perform multimodal learning analytics. Our first step is to align the pre-processed data from multiple sensors. Then analyze the interaction between the learner’s performance and emotions. In our recent study, we collected eye-gaze patterns, facial emotions, and log data from 18 students working on an OELE called Betty’s Brain. We are pre-processing the data from multiple sensors. The preliminary results of these works are published in Educational Data Mining, Intelligent Tutoring Systems, and UMAP Conferences.
Open-ended computer-based learning environments (OELEs) focus on developing students’ a) cognitive skills, b) metacognitive processes and c) problem-solving strategies that go beyond the acquisition of domain-specific cognitive skills. However, novice students often have difficulties in making progress when working on complex problems in OELE. To help such students with personalized and adaptive feedback, I developed a learner-modeling scheme that captures both the learner’s performance and their learning behaviors as they interact with the system. Learning behaviors combine the learner’s use of cognitive skills and their problem-solving strategies. The learner-modeling framework is developed by extending the framework in Generalized Intelligent Framework for Tutoring (GIFT), developed by U.S. Army Research Laboratory. The learner-modeling framework has been applied to an OELE called UrbanSim, a turn-based simulation environment for counterinsurgency (COIN) training of army officers, who play the role of a commander directing counterinsurgency efforts in a Middle Eastern region. The results of this work are published in IEEE Trans on Learning Technologies. Recently we used Process Mining to model learners' behavior in OELEs and the results are published in ICCE 2018 conference.
I have developed an algorithm to extract meaningful answers from scattered information including verbal interactions from the online discussion forums. First, we pre-processed the data from online discussion forums (ODF) to classify posts as questions, answers, additional questions, and probing questions. Then we developed an algorithm to combine multiple pre-processed sentences based on context, and problem to generate answers automatically. This work has been submitted for a patent: WO2017119014A1.
We developed a system that would support humans in making decisions by processing a large amount of text data and creating hidden hypotheses from the data. To develop the automated reasoning system, I developed an algorithm to extract and detect temporal relations between events from large text data, such as news corpora. However, news corpora had a few temporal relations hence to overcome this problem, word embeddings are used, in particular, neural networks architectures to exploit word embeddings for temporal relation detection. We developed an automated reasoning engine using deep learning networks for reasoning in an embedded space. The Results of this work are published in a workshop at NIPS and AI Communications.
I developed a theory-based data-driven (theory-driven) model to detect frustration. The model is applied to data collected from 27 sixth-grade students. The results with details of the frustration model were published in ICALT 2012 and IEEE Trans in Learning Technology, 2013. We developed an algorithm to respond to the student’s frustration with motivational messages during their interaction with ITS. The approach was tested in three schools in India, and the number of frustration instances per session after implementing the algorithm was analyzed. The frustration instances were reduced statistically significantly (P < 0.05), due to the motivational messages. The results are published in IEEE Trans of Learning Technologies.